Allow top-level string literals in Core (#8472)
[ghc.git] / compiler / simplCore / Simplify.hs
1 {-
2 (c) The AQUA Project, Glasgow University, 1993-1998
3
4 \section[Simplify]{The main module of the simplifier}
5 -}
6
7 {-# LANGUAGE CPP #-}
8
9 module Simplify ( simplTopBinds, simplExpr, simplRules ) where
10
11 #include "HsVersions.h"
12
13 import DynFlags
14 import SimplMonad
15 import Type hiding ( substTy, substTyVar, extendTvSubst, extendCvSubst )
16 import SimplEnv
17 import SimplUtils
18 import FamInstEnv ( FamInstEnv )
19 import Literal ( litIsLifted ) --, mkMachInt ) -- temporalily commented out. See #8326
20 import Id
21 import MkId ( seqId, voidPrimId )
22 import MkCore ( mkImpossibleExpr, castBottomExpr )
23 import IdInfo
24 import Name ( Name, mkSystemVarName, isExternalName, getOccFS )
25 import Coercion hiding ( substCo, substCoVar )
26 import OptCoercion ( optCoercion )
27 import FamInstEnv ( topNormaliseType_maybe )
28 import DataCon ( DataCon, dataConWorkId, dataConRepStrictness
29 , isMarkedStrict, dataConRepArgTys ) --, dataConTyCon, dataConTag, fIRST_TAG )
30 --import TyCon ( isEnumerationTyCon ) -- temporalily commented out. See #8326
31 import CoreMonad ( Tick(..), SimplifierMode(..) )
32 import CoreSyn
33 import Demand ( StrictSig(..), dmdTypeDepth, isStrictDmd )
34 import PprCore ( pprCoreExpr )
35 import CoreUnfold
36 import CoreUtils
37 import CoreArity
38 import CoreSubst ( pushCoTyArg, pushCoValArg )
39 --import PrimOp ( tagToEnumKey ) -- temporalily commented out. See #8326
40 import Rules ( mkRuleInfo, lookupRule, getRules )
41 import TysPrim ( voidPrimTy ) --, intPrimTy ) -- temporalily commented out. See #8326
42 import BasicTypes ( TopLevelFlag(..), isTopLevel, RecFlag(..) )
43 import MonadUtils ( foldlM, mapAccumLM, liftIO )
44 import Maybes ( orElse )
45 --import Unique ( hasKey ) -- temporalily commented out. See #8326
46 import Control.Monad
47 import Outputable
48 import FastString
49 import Pair
50 import Util
51 import ErrUtils
52
53 {-
54 The guts of the simplifier is in this module, but the driver loop for
55 the simplifier is in SimplCore.hs.
56
57
58 -----------------------------------------
59 *** IMPORTANT NOTE ***
60 -----------------------------------------
61 The simplifier used to guarantee that the output had no shadowing, but
62 it does not do so any more. (Actually, it never did!) The reason is
63 documented with simplifyArgs.
64
65
66 -----------------------------------------
67 *** IMPORTANT NOTE ***
68 -----------------------------------------
69 Many parts of the simplifier return a bunch of "floats" as well as an
70 expression. This is wrapped as a datatype SimplUtils.FloatsWith.
71
72 All "floats" are let-binds, not case-binds, but some non-rec lets may
73 be unlifted (with RHS ok-for-speculation).
74
75
76
77 -----------------------------------------
78 ORGANISATION OF FUNCTIONS
79 -----------------------------------------
80 simplTopBinds
81 - simplify all top-level binders
82 - for NonRec, call simplRecOrTopPair
83 - for Rec, call simplRecBind
84
85
86 ------------------------------
87 simplExpr (applied lambda) ==> simplNonRecBind
88 simplExpr (Let (NonRec ...) ..) ==> simplNonRecBind
89 simplExpr (Let (Rec ...) ..) ==> simplify binders; simplRecBind
90
91 ------------------------------
92 simplRecBind [binders already simplfied]
93 - use simplRecOrTopPair on each pair in turn
94
95 simplRecOrTopPair [binder already simplified]
96 Used for: recursive bindings (top level and nested)
97 top-level non-recursive bindings
98 Returns:
99 - check for PreInlineUnconditionally
100 - simplLazyBind
101
102 simplNonRecBind
103 Used for: non-top-level non-recursive bindings
104 beta reductions (which amount to the same thing)
105 Because it can deal with strict arts, it takes a
106 "thing-inside" and returns an expression
107
108 - check for PreInlineUnconditionally
109 - simplify binder, including its IdInfo
110 - if strict binding
111 simplStrictArg
112 mkAtomicArgs
113 completeNonRecX
114 else
115 simplLazyBind
116 addFloats
117
118 simplNonRecX: [given a *simplified* RHS, but an *unsimplified* binder]
119 Used for: binding case-binder and constr args in a known-constructor case
120 - check for PreInLineUnconditionally
121 - simplify binder
122 - completeNonRecX
123
124 ------------------------------
125 simplLazyBind: [binder already simplified, RHS not]
126 Used for: recursive bindings (top level and nested)
127 top-level non-recursive bindings
128 non-top-level, but *lazy* non-recursive bindings
129 [must not be strict or unboxed]
130 Returns floats + an augmented environment, not an expression
131 - substituteIdInfo and add result to in-scope
132 [so that rules are available in rec rhs]
133 - simplify rhs
134 - mkAtomicArgs
135 - float if exposes constructor or PAP
136 - completeBind
137
138
139 completeNonRecX: [binder and rhs both simplified]
140 - if the the thing needs case binding (unlifted and not ok-for-spec)
141 build a Case
142 else
143 completeBind
144 addFloats
145
146 completeBind: [given a simplified RHS]
147 [used for both rec and non-rec bindings, top level and not]
148 - try PostInlineUnconditionally
149 - add unfolding [this is the only place we add an unfolding]
150 - add arity
151
152
153
154 Right hand sides and arguments
155 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
156 In many ways we want to treat
157 (a) the right hand side of a let(rec), and
158 (b) a function argument
159 in the same way. But not always! In particular, we would
160 like to leave these arguments exactly as they are, so they
161 will match a RULE more easily.
162
163 f (g x, h x)
164 g (+ x)
165
166 It's harder to make the rule match if we ANF-ise the constructor,
167 or eta-expand the PAP:
168
169 f (let { a = g x; b = h x } in (a,b))
170 g (\y. + x y)
171
172 On the other hand if we see the let-defns
173
174 p = (g x, h x)
175 q = + x
176
177 then we *do* want to ANF-ise and eta-expand, so that p and q
178 can be safely inlined.
179
180 Even floating lets out is a bit dubious. For let RHS's we float lets
181 out if that exposes a value, so that the value can be inlined more vigorously.
182 For example
183
184 r = let x = e in (x,x)
185
186 Here, if we float the let out we'll expose a nice constructor. We did experiments
187 that showed this to be a generally good thing. But it was a bad thing to float
188 lets out unconditionally, because that meant they got allocated more often.
189
190 For function arguments, there's less reason to expose a constructor (it won't
191 get inlined). Just possibly it might make a rule match, but I'm pretty skeptical.
192 So for the moment we don't float lets out of function arguments either.
193
194
195 Eta expansion
196 ~~~~~~~~~~~~~~
197 For eta expansion, we want to catch things like
198
199 case e of (a,b) -> \x -> case a of (p,q) -> \y -> r
200
201 If the \x was on the RHS of a let, we'd eta expand to bring the two
202 lambdas together. And in general that's a good thing to do. Perhaps
203 we should eta expand wherever we find a (value) lambda? Then the eta
204 expansion at a let RHS can concentrate solely on the PAP case.
205
206
207 ************************************************************************
208 * *
209 \subsection{Bindings}
210 * *
211 ************************************************************************
212 -}
213
214 simplTopBinds :: SimplEnv -> [InBind] -> SimplM SimplEnv
215
216 simplTopBinds env0 binds0
217 = do { -- Put all the top-level binders into scope at the start
218 -- so that if a transformation rule has unexpectedly brought
219 -- anything into scope, then we don't get a complaint about that.
220 -- It's rather as if the top-level binders were imported.
221 -- See note [Glomming] in OccurAnal.
222 ; env1 <- simplRecBndrs env0 (bindersOfBinds binds0)
223 ; env2 <- simpl_binds env1 binds0
224 ; freeTick SimplifierDone
225 ; return env2 }
226 where
227 -- We need to track the zapped top-level binders, because
228 -- they should have their fragile IdInfo zapped (notably occurrence info)
229 -- That's why we run down binds and bndrs' simultaneously.
230 --
231 simpl_binds :: SimplEnv -> [InBind] -> SimplM SimplEnv
232 simpl_binds env [] = return env
233 simpl_binds env (bind:binds) = do { env' <- simpl_bind env bind
234 ; simpl_binds env' binds }
235
236 simpl_bind env (Rec pairs) = simplRecBind env TopLevel pairs
237 simpl_bind env (NonRec b r) = do { (env', b') <- addBndrRules env b (lookupRecBndr env b)
238 ; simplRecOrTopPair env' TopLevel NonRecursive b b' r }
239
240 {-
241 ************************************************************************
242 * *
243 \subsection{Lazy bindings}
244 * *
245 ************************************************************************
246
247 simplRecBind is used for
248 * recursive bindings only
249 -}
250
251 simplRecBind :: SimplEnv -> TopLevelFlag
252 -> [(InId, InExpr)]
253 -> SimplM SimplEnv
254 simplRecBind env0 top_lvl pairs0
255 = do { (env_with_info, triples) <- mapAccumLM add_rules env0 pairs0
256 ; env1 <- go (zapFloats env_with_info) triples
257 ; return (env0 `addRecFloats` env1) }
258 -- addFloats adds the floats from env1,
259 -- _and_ updates env0 with the in-scope set from env1
260 where
261 add_rules :: SimplEnv -> (InBndr,InExpr) -> SimplM (SimplEnv, (InBndr, OutBndr, InExpr))
262 -- Add the (substituted) rules to the binder
263 add_rules env (bndr, rhs)
264 = do { (env', bndr') <- addBndrRules env bndr (lookupRecBndr env bndr)
265 ; return (env', (bndr, bndr', rhs)) }
266
267 go env [] = return env
268
269 go env ((old_bndr, new_bndr, rhs) : pairs)
270 = do { env' <- simplRecOrTopPair env top_lvl Recursive old_bndr new_bndr rhs
271 ; go env' pairs }
272
273 {-
274 simplOrTopPair is used for
275 * recursive bindings (whether top level or not)
276 * top-level non-recursive bindings
277
278 It assumes the binder has already been simplified, but not its IdInfo.
279 -}
280
281 simplRecOrTopPair :: SimplEnv
282 -> TopLevelFlag -> RecFlag
283 -> InId -> OutBndr -> InExpr -- Binder and rhs
284 -> SimplM SimplEnv -- Returns an env that includes the binding
285
286 simplRecOrTopPair env top_lvl is_rec old_bndr new_bndr rhs
287 = do { dflags <- getDynFlags
288 ; trace_bind dflags $
289 if preInlineUnconditionally dflags env top_lvl old_bndr rhs
290 -- Check for unconditional inline
291 then do tick (PreInlineUnconditionally old_bndr)
292 return (extendIdSubst env old_bndr (mkContEx env rhs))
293 else simplLazyBind env top_lvl is_rec old_bndr new_bndr rhs env }
294 where
295 trace_bind dflags thing_inside
296 | not (dopt Opt_D_verbose_core2core dflags)
297 = thing_inside
298 | otherwise
299 = pprTrace "SimplBind" (ppr old_bndr) thing_inside
300 -- trace_bind emits a trace for each top-level binding, which
301 -- helps to locate the tracing for inlining and rule firing
302
303 {-
304 simplLazyBind is used for
305 * [simplRecOrTopPair] recursive bindings (whether top level or not)
306 * [simplRecOrTopPair] top-level non-recursive bindings
307 * [simplNonRecE] non-top-level *lazy* non-recursive bindings
308
309 Nota bene:
310 1. It assumes that the binder is *already* simplified,
311 and is in scope, and its IdInfo too, except unfolding
312
313 2. It assumes that the binder type is lifted.
314
315 3. It does not check for pre-inline-unconditionally;
316 that should have been done already.
317 -}
318
319 simplLazyBind :: SimplEnv
320 -> TopLevelFlag -> RecFlag
321 -> InId -> OutId -- Binder, both pre-and post simpl
322 -- The OutId has IdInfo, except arity, unfolding
323 -> InExpr -> SimplEnv -- The RHS and its environment
324 -> SimplM SimplEnv
325 -- Precondition: rhs obeys the let/app invariant
326 simplLazyBind env top_lvl is_rec bndr bndr1 rhs rhs_se
327 = -- pprTrace "simplLazyBind" ((ppr bndr <+> ppr bndr1) $$ ppr rhs $$ ppr (seIdSubst rhs_se)) $
328 do { let rhs_env = rhs_se `setInScope` env
329 (tvs, body) = case collectTyAndValBinders rhs of
330 (tvs, [], body)
331 | surely_not_lam body -> (tvs, body)
332 _ -> ([], rhs)
333
334 surely_not_lam (Lam {}) = False
335 surely_not_lam (Tick t e)
336 | not (tickishFloatable t) = surely_not_lam e
337 -- eta-reduction could float
338 surely_not_lam _ = True
339 -- Do not do the "abstract tyyvar" thing if there's
340 -- a lambda inside, because it defeats eta-reduction
341 -- f = /\a. \x. g a x
342 -- should eta-reduce.
343
344
345 ; (body_env, tvs') <- simplBinders rhs_env tvs
346 -- See Note [Floating and type abstraction] in SimplUtils
347
348 -- Simplify the RHS
349 ; let rhs_cont = mkRhsStop (substTy body_env (exprType body))
350 ; (body_env1, body1) <- simplExprF body_env body rhs_cont
351 -- ANF-ise a constructor or PAP rhs
352 ; (body_env2, body2) <- prepareRhs top_lvl body_env1 bndr1 body1
353
354 ; (env', rhs')
355 <- if not (doFloatFromRhs top_lvl is_rec False body2 body_env2)
356 then -- No floating, revert to body1
357 do { rhs' <- mkLam env tvs' (wrapFloats body_env1 body1) rhs_cont
358 ; return (env, rhs') }
359
360 else if null tvs then -- Simple floating
361 do { tick LetFloatFromLet
362 ; return (addFloats env body_env2, body2) }
363
364 else -- Do type-abstraction first
365 do { tick LetFloatFromLet
366 ; (poly_binds, body3) <- abstractFloats tvs' body_env2 body2
367 ; rhs' <- mkLam env tvs' body3 rhs_cont
368 ; env' <- foldlM (addPolyBind top_lvl) env poly_binds
369 ; return (env', rhs') }
370
371 ; completeBind env' top_lvl bndr bndr1 rhs' }
372
373 {-
374 A specialised variant of simplNonRec used when the RHS is already simplified,
375 notably in knownCon. It uses case-binding where necessary.
376 -}
377
378 simplNonRecX :: SimplEnv
379 -> InId -- Old binder
380 -> OutExpr -- Simplified RHS
381 -> SimplM SimplEnv
382 -- Precondition: rhs satisfies the let/app invariant
383 simplNonRecX env bndr new_rhs
384 | isDeadBinder bndr -- Not uncommon; e.g. case (a,b) of c { (p,q) -> p }
385 = return env -- Here c is dead, and we avoid creating
386 -- the binding c = (a,b)
387
388 | Coercion co <- new_rhs
389 = return (extendCvSubst env bndr co)
390
391 | otherwise
392 = do { (env', bndr') <- simplBinder env bndr
393 ; completeNonRecX NotTopLevel env' (isStrictId bndr) bndr bndr' new_rhs }
394 -- simplNonRecX is only used for NotTopLevel things
395
396 completeNonRecX :: TopLevelFlag -> SimplEnv
397 -> Bool
398 -> InId -- Old binder
399 -> OutId -- New binder
400 -> OutExpr -- Simplified RHS
401 -> SimplM SimplEnv
402 -- Precondition: rhs satisfies the let/app invariant
403 -- See Note [CoreSyn let/app invariant] in CoreSyn
404
405 completeNonRecX top_lvl env is_strict old_bndr new_bndr new_rhs
406 = do { (env1, rhs1) <- prepareRhs top_lvl (zapFloats env) new_bndr new_rhs
407 ; (env2, rhs2) <-
408 if doFloatFromRhs NotTopLevel NonRecursive is_strict rhs1 env1
409 then do { tick LetFloatFromLet
410 ; return (addFloats env env1, rhs1) } -- Add the floats to the main env
411 else return (env, wrapFloats env1 rhs1) -- Wrap the floats around the RHS
412 ; completeBind env2 NotTopLevel old_bndr new_bndr rhs2 }
413
414 {-
415 {- No, no, no! Do not try preInlineUnconditionally in completeNonRecX
416 Doing so risks exponential behaviour, because new_rhs has been simplified once already
417 In the cases described by the folowing commment, postInlineUnconditionally will
418 catch many of the relevant cases.
419 -- This happens; for example, the case_bndr during case of
420 -- known constructor: case (a,b) of x { (p,q) -> ... }
421 -- Here x isn't mentioned in the RHS, so we don't want to
422 -- create the (dead) let-binding let x = (a,b) in ...
423 --
424 -- Similarly, single occurrences can be inlined vigourously
425 -- e.g. case (f x, g y) of (a,b) -> ....
426 -- If a,b occur once we can avoid constructing the let binding for them.
427
428 Furthermore in the case-binding case preInlineUnconditionally risks extra thunks
429 -- Consider case I# (quotInt# x y) of
430 -- I# v -> let w = J# v in ...
431 -- If we gaily inline (quotInt# x y) for v, we end up building an
432 -- extra thunk:
433 -- let w = J# (quotInt# x y) in ...
434 -- because quotInt# can fail.
435
436 | preInlineUnconditionally env NotTopLevel bndr new_rhs
437 = thing_inside (extendIdSubst env bndr (DoneEx new_rhs))
438 -}
439
440 ----------------------------------
441 prepareRhs takes a putative RHS, checks whether it's a PAP or
442 constructor application and, if so, converts it to ANF, so that the
443 resulting thing can be inlined more easily. Thus
444 x = (f a, g b)
445 becomes
446 t1 = f a
447 t2 = g b
448 x = (t1,t2)
449
450 We also want to deal well cases like this
451 v = (f e1 `cast` co) e2
452 Here we want to make e1,e2 trivial and get
453 x1 = e1; x2 = e2; v = (f x1 `cast` co) v2
454 That's what the 'go' loop in prepareRhs does
455 -}
456
457 prepareRhs :: TopLevelFlag -> SimplEnv -> OutId -> OutExpr -> SimplM (SimplEnv, OutExpr)
458 -- Adds new floats to the env iff that allows us to return a good RHS
459 prepareRhs top_lvl env id (Cast rhs co) -- Note [Float coercions]
460 | Pair ty1 _ty2 <- coercionKind co -- Do *not* do this if rhs has an unlifted type
461 , not (isUnliftedType ty1) -- see Note [Float coercions (unlifted)]
462 = do { (env', rhs') <- makeTrivialWithInfo top_lvl env (getOccFS id) sanitised_info rhs
463 ; return (env', Cast rhs' co) }
464 where
465 sanitised_info = vanillaIdInfo `setStrictnessInfo` strictnessInfo info
466 `setDemandInfo` demandInfo info
467 info = idInfo id
468
469 prepareRhs top_lvl env0 id rhs0
470 = do { (_is_exp, env1, rhs1) <- go 0 env0 rhs0
471 ; return (env1, rhs1) }
472 where
473 go n_val_args env (Cast rhs co)
474 = do { (is_exp, env', rhs') <- go n_val_args env rhs
475 ; return (is_exp, env', Cast rhs' co) }
476 go n_val_args env (App fun (Type ty))
477 = do { (is_exp, env', rhs') <- go n_val_args env fun
478 ; return (is_exp, env', App rhs' (Type ty)) }
479 go n_val_args env (App fun arg)
480 = do { (is_exp, env', fun') <- go (n_val_args+1) env fun
481 ; case is_exp of
482 True -> do { (env'', arg') <- makeTrivial top_lvl env' (getOccFS id) arg
483 ; return (True, env'', App fun' arg') }
484 False -> return (False, env, App fun arg) }
485 go n_val_args env (Var fun)
486 = return (is_exp, env, Var fun)
487 where
488 is_exp = isExpandableApp fun n_val_args -- The fun a constructor or PAP
489 -- See Note [CONLIKE pragma] in BasicTypes
490 -- The definition of is_exp should match that in
491 -- OccurAnal.occAnalApp
492
493 go n_val_args env (Tick t rhs)
494 -- We want to be able to float bindings past this
495 -- tick. Non-scoping ticks don't care.
496 | tickishScoped t == NoScope
497 = do { (is_exp, env', rhs') <- go n_val_args env rhs
498 ; return (is_exp, env', Tick t rhs') }
499 -- On the other hand, for scoping ticks we need to be able to
500 -- copy them on the floats, which in turn is only allowed if
501 -- we can obtain non-counting ticks.
502 | not (tickishCounts t) || tickishCanSplit t
503 = do { (is_exp, env', rhs') <- go n_val_args (zapFloats env) rhs
504 ; let tickIt (id, expr) = (id, mkTick (mkNoCount t) expr)
505 floats' = seFloats $ env `addFloats` mapFloats env' tickIt
506 ; return (is_exp, env' { seFloats = floats' }, Tick t rhs') }
507
508 go _ env other
509 = return (False, env, other)
510
511 {-
512 Note [Float coercions]
513 ~~~~~~~~~~~~~~~~~~~~~~
514 When we find the binding
515 x = e `cast` co
516 we'd like to transform it to
517 x' = e
518 x = x `cast` co -- A trivial binding
519 There's a chance that e will be a constructor application or function, or something
520 like that, so moving the coercion to the usage site may well cancel the coercions
521 and lead to further optimisation. Example:
522
523 data family T a :: *
524 data instance T Int = T Int
525
526 foo :: Int -> Int -> Int
527 foo m n = ...
528 where
529 x = T m
530 go 0 = 0
531 go n = case x of { T m -> go (n-m) }
532 -- This case should optimise
533
534 Note [Preserve strictness when floating coercions]
535 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
536 In the Note [Float coercions] transformation, keep the strictness info.
537 Eg
538 f = e `cast` co -- f has strictness SSL
539 When we transform to
540 f' = e -- f' also has strictness SSL
541 f = f' `cast` co -- f still has strictness SSL
542
543 Its not wrong to drop it on the floor, but better to keep it.
544
545 Note [Float coercions (unlifted)]
546 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
547 BUT don't do [Float coercions] if 'e' has an unlifted type.
548 This *can* happen:
549
550 foo :: Int = (error (# Int,Int #) "urk")
551 `cast` CoUnsafe (# Int,Int #) Int
552
553 If do the makeTrivial thing to the error call, we'll get
554 foo = case error (# Int,Int #) "urk" of v -> v `cast` ...
555 But 'v' isn't in scope!
556
557 These strange casts can happen as a result of case-of-case
558 bar = case (case x of { T -> (# 2,3 #); F -> error "urk" }) of
559 (# p,q #) -> p+q
560 -}
561
562 makeTrivialArg :: SimplEnv -> ArgSpec -> SimplM (SimplEnv, ArgSpec)
563 makeTrivialArg env (ValArg e) = do
564 { (env', e') <- makeTrivial NotTopLevel env (fsLit "arg") e
565 ; return (env', ValArg e') }
566 makeTrivialArg env arg = return (env, arg) -- CastBy, TyArg
567
568 makeTrivial :: TopLevelFlag -> SimplEnv
569 -> FastString -- ^ a "friendly name" to build the new binder from
570 -> OutExpr -> SimplM (SimplEnv, OutExpr)
571 -- Binds the expression to a variable, if it's not trivial, returning the variable
572 makeTrivial top_lvl env context expr =
573 makeTrivialWithInfo top_lvl env context vanillaIdInfo expr
574
575 makeTrivialWithInfo :: TopLevelFlag -> SimplEnv
576 -> FastString
577 -- ^ a "friendly name" to build the new binder from
578 -> IdInfo -> OutExpr -> SimplM (SimplEnv, OutExpr)
579 -- Propagate strictness and demand info to the new binder
580 -- Note [Preserve strictness when floating coercions]
581 -- Returned SimplEnv has same substitution as incoming one
582 makeTrivialWithInfo top_lvl env context info expr
583 | exprIsTrivial expr -- Already trivial
584 || not (bindingOk top_lvl expr) -- Cannot trivialise
585 -- See Note [Cannot trivialise]
586 = return (env, expr)
587 | otherwise -- See Note [Take care] below
588 = do { uniq <- getUniqueM
589 ; let name = mkSystemVarName uniq context
590 var = mkLocalIdOrCoVarWithInfo name expr_ty info
591 ; env' <- completeNonRecX top_lvl env False var var expr
592 ; expr' <- simplVar env' var
593 ; return (env', expr') }
594 -- The simplVar is needed because we're constructing a new binding
595 -- a = rhs
596 -- And if rhs is of form (rhs1 |> co), then we might get
597 -- a1 = rhs1
598 -- a = a1 |> co
599 -- and now a's RHS is trivial and can be substituted out, and that
600 -- is what completeNonRecX will do
601 -- To put it another way, it's as if we'd simplified
602 -- let var = e in var
603 where
604 expr_ty = exprType expr
605
606 bindingOk :: TopLevelFlag -> CoreExpr -> Bool
607 -- True iff we can have a binding of this expression at this level
608 -- Precondition: the type is the type of the expression
609 bindingOk top_lvl expr
610 | isTopLevel top_lvl = exprIsTopLevelBindable expr
611 | otherwise = True
612
613 {-
614 Note [Cannot trivialise]
615 ~~~~~~~~~~~~~~~~~~~~~~~~
616 Consider tih
617 f :: Int -> Addr#
618
619 foo :: Bar
620 foo = Bar (f 3)
621
622 Then we can't ANF-ise foo, even though we'd like to, because
623 we can't make a top-level binding for the Addr# (f 3). And if
624 so we don't want to turn it into
625 foo = let x = f 3 in Bar x
626 because we'll just end up inlining x back, and that makes the
627 simplifier loop. Better not to ANF-ise it at all.
628
629 Literal strings are an exception.
630
631 foo = Ptr "blob"#
632
633 We want to turn this into:
634
635 foo1 = "blob"#
636 foo = Ptr foo1
637
638 See Note [CoreSyn top-level string literals] in CoreSyn.
639
640 ************************************************************************
641 * *
642 \subsection{Completing a lazy binding}
643 * *
644 ************************************************************************
645
646 completeBind
647 * deals only with Ids, not TyVars
648 * takes an already-simplified binder and RHS
649 * is used for both recursive and non-recursive bindings
650 * is used for both top-level and non-top-level bindings
651
652 It does the following:
653 - tries discarding a dead binding
654 - tries PostInlineUnconditionally
655 - add unfolding [this is the only place we add an unfolding]
656 - add arity
657
658 It does *not* attempt to do let-to-case. Why? Because it is used for
659 - top-level bindings (when let-to-case is impossible)
660 - many situations where the "rhs" is known to be a WHNF
661 (so let-to-case is inappropriate).
662
663 Nor does it do the atomic-argument thing
664 -}
665
666 completeBind :: SimplEnv
667 -> TopLevelFlag -- Flag stuck into unfolding
668 -> InId -- Old binder
669 -> OutId -> OutExpr -- New binder and RHS
670 -> SimplM SimplEnv
671 -- completeBind may choose to do its work
672 -- * by extending the substitution (e.g. let x = y in ...)
673 -- * or by adding to the floats in the envt
674 --
675 -- Precondition: rhs obeys the let/app invariant
676 completeBind env top_lvl old_bndr new_bndr new_rhs
677 | isCoVar old_bndr
678 = case new_rhs of
679 Coercion co -> return (extendCvSubst env old_bndr co)
680 _ -> return (addNonRec env new_bndr new_rhs)
681
682 | otherwise
683 = ASSERT( isId new_bndr )
684 do { let old_info = idInfo old_bndr
685 old_unf = unfoldingInfo old_info
686 occ_info = occInfo old_info
687
688 -- Do eta-expansion on the RHS of the binding
689 -- See Note [Eta-expanding at let bindings] in SimplUtils
690 ; (new_arity, final_rhs) <- tryEtaExpandRhs env new_bndr new_rhs
691
692 -- Simplify the unfolding
693 ; new_unfolding <- simplLetUnfolding env top_lvl old_bndr final_rhs old_unf
694
695 ; dflags <- getDynFlags
696 ; if postInlineUnconditionally dflags env top_lvl new_bndr occ_info
697 final_rhs new_unfolding
698
699 -- Inline and discard the binding
700 then do { tick (PostInlineUnconditionally old_bndr)
701 ; return (extendIdSubst env old_bndr (DoneEx final_rhs)) }
702 -- Use the substitution to make quite, quite sure that the
703 -- substitution will happen, since we are going to discard the binding
704 else
705 do { let info1 = idInfo new_bndr `setArityInfo` new_arity
706
707 -- Unfolding info: Note [Setting the new unfolding]
708 info2 = info1 `setUnfoldingInfo` new_unfolding
709
710 -- Demand info: Note [Setting the demand info]
711 --
712 -- We also have to nuke demand info if for some reason
713 -- eta-expansion *reduces* the arity of the binding to less
714 -- than that of the strictness sig. This can happen: see Note [Arity decrease].
715 info3 | isEvaldUnfolding new_unfolding
716 || (case strictnessInfo info2 of
717 StrictSig dmd_ty -> new_arity < dmdTypeDepth dmd_ty)
718 = zapDemandInfo info2 `orElse` info2
719 | otherwise
720 = info2
721
722 final_id = new_bndr `setIdInfo` info3
723
724 ; -- pprTrace "Binding" (ppr final_id <+> ppr new_unfolding) $
725 return (addNonRec env final_id final_rhs) } }
726 -- The addNonRec adds it to the in-scope set too
727
728 ------------------------------
729 addPolyBind :: TopLevelFlag -> SimplEnv -> OutBind -> SimplM SimplEnv
730 -- Add a new binding to the environment, complete with its unfolding
731 -- but *do not* do postInlineUnconditionally, because we have already
732 -- processed some of the scope of the binding
733 -- We still want the unfolding though. Consider
734 -- let
735 -- x = /\a. let y = ... in Just y
736 -- in body
737 -- Then we float the y-binding out (via abstractFloats and addPolyBind)
738 -- but 'x' may well then be inlined in 'body' in which case we'd like the
739 -- opportunity to inline 'y' too.
740 --
741 -- INVARIANT: the arity is correct on the incoming binders
742
743 addPolyBind top_lvl env (NonRec poly_id rhs)
744 = do { unfolding <- simplLetUnfolding env top_lvl poly_id rhs noUnfolding
745 -- Assumes that poly_id did not have an INLINE prag
746 -- which is perhaps wrong. ToDo: think about this
747 ; let final_id = setIdInfo poly_id $
748 idInfo poly_id `setUnfoldingInfo` unfolding
749
750 ; return (addNonRec env final_id rhs) }
751
752 addPolyBind _ env bind@(Rec _)
753 = return (extendFloats env bind)
754 -- Hack: letrecs are more awkward, so we extend "by steam"
755 -- without adding unfoldings etc. At worst this leads to
756 -- more simplifier iterations
757
758 {- Note [Arity decrease]
759 ~~~~~~~~~~~~~~~~~~~~~~~~
760 Generally speaking the arity of a binding should not decrease. But it *can*
761 legitimately happen because of RULES. Eg
762 f = g Int
763 where g has arity 2, will have arity 2. But if there's a rewrite rule
764 g Int --> h
765 where h has arity 1, then f's arity will decrease. Here's a real-life example,
766 which is in the output of Specialise:
767
768 Rec {
769 $dm {Arity 2} = \d.\x. op d
770 {-# RULES forall d. $dm Int d = $s$dm #-}
771
772 dInt = MkD .... opInt ...
773 opInt {Arity 1} = $dm dInt
774
775 $s$dm {Arity 0} = \x. op dInt }
776
777 Here opInt has arity 1; but when we apply the rule its arity drops to 0.
778 That's why Specialise goes to a little trouble to pin the right arity
779 on specialised functions too.
780
781 Note [Setting the demand info]
782 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
783 If the unfolding is a value, the demand info may
784 go pear-shaped, so we nuke it. Example:
785 let x = (a,b) in
786 case x of (p,q) -> h p q x
787 Here x is certainly demanded. But after we've nuked
788 the case, we'll get just
789 let x = (a,b) in h a b x
790 and now x is not demanded (I'm assuming h is lazy)
791 This really happens. Similarly
792 let f = \x -> e in ...f..f...
793 After inlining f at some of its call sites the original binding may
794 (for example) be no longer strictly demanded.
795 The solution here is a bit ad hoc...
796
797
798 ************************************************************************
799 * *
800 \subsection[Simplify-simplExpr]{The main function: simplExpr}
801 * *
802 ************************************************************************
803
804 The reason for this OutExprStuff stuff is that we want to float *after*
805 simplifying a RHS, not before. If we do so naively we get quadratic
806 behaviour as things float out.
807
808 To see why it's important to do it after, consider this (real) example:
809
810 let t = f x
811 in fst t
812 ==>
813 let t = let a = e1
814 b = e2
815 in (a,b)
816 in fst t
817 ==>
818 let a = e1
819 b = e2
820 t = (a,b)
821 in
822 a -- Can't inline a this round, cos it appears twice
823 ==>
824 e1
825
826 Each of the ==> steps is a round of simplification. We'd save a
827 whole round if we float first. This can cascade. Consider
828
829 let f = g d
830 in \x -> ...f...
831 ==>
832 let f = let d1 = ..d.. in \y -> e
833 in \x -> ...f...
834 ==>
835 let d1 = ..d..
836 in \x -> ...(\y ->e)...
837
838 Only in this second round can the \y be applied, and it
839 might do the same again.
840 -}
841
842 simplExpr :: SimplEnv -> CoreExpr -> SimplM CoreExpr
843 simplExpr env expr = simplExprC env expr (mkBoringStop expr_out_ty)
844 where
845 expr_out_ty :: OutType
846 expr_out_ty = substTy env (exprType expr)
847
848 simplExprC :: SimplEnv -> CoreExpr -> SimplCont -> SimplM CoreExpr
849 -- Simplify an expression, given a continuation
850 simplExprC env expr cont
851 = -- pprTrace "simplExprC" (ppr expr $$ ppr cont {- $$ ppr (seIdSubst env) -} $$ ppr (seFloats env) ) $
852 do { (env', expr') <- simplExprF (zapFloats env) expr cont
853 ; -- pprTrace "simplExprC ret" (ppr expr $$ ppr expr') $
854 -- pprTrace "simplExprC ret3" (ppr (seInScope env')) $
855 -- pprTrace "simplExprC ret4" (ppr (seFloats env')) $
856 return (wrapFloats env' expr') }
857
858 --------------------------------------------------
859 simplExprF :: SimplEnv -> InExpr -> SimplCont
860 -> SimplM (SimplEnv, OutExpr)
861
862 simplExprF env e cont
863 = {- pprTrace "simplExprF" (vcat
864 [ ppr e
865 , text "cont =" <+> ppr cont
866 , text "inscope =" <+> ppr (seInScope env)
867 , text "tvsubst =" <+> ppr (seTvSubst env)
868 , text "idsubst =" <+> ppr (seIdSubst env)
869 , text "cvsubst =" <+> ppr (seCvSubst env)
870 {- , ppr (seFloats env) -}
871 ]) $ -}
872 simplExprF1 env e cont
873
874 simplExprF1 :: SimplEnv -> InExpr -> SimplCont
875 -> SimplM (SimplEnv, OutExpr)
876 simplExprF1 env (Var v) cont = simplIdF env v cont
877 simplExprF1 env (Lit lit) cont = rebuild env (Lit lit) cont
878 simplExprF1 env (Tick t expr) cont = simplTick env t expr cont
879 simplExprF1 env (Cast body co) cont = simplCast env body co cont
880 simplExprF1 env (Coercion co) cont = simplCoercionF env co cont
881 simplExprF1 env (Type ty) cont = ASSERT( contIsRhsOrArg cont )
882 rebuild env (Type (substTy env ty)) cont
883
884 simplExprF1 env (App fun arg) cont
885 = simplExprF env fun $
886 case arg of
887 Type ty -> ApplyToTy { sc_arg_ty = substTy env ty
888 , sc_hole_ty = substTy env (exprType fun)
889 , sc_cont = cont }
890 _ -> ApplyToVal { sc_arg = arg, sc_env = env
891 , sc_dup = NoDup, sc_cont = cont }
892
893 simplExprF1 env expr@(Lam {}) cont
894 = simplLam env zapped_bndrs body cont
895 -- The main issue here is under-saturated lambdas
896 -- (\x1. \x2. e) arg1
897 -- Here x1 might have "occurs-once" occ-info, because occ-info
898 -- is computed assuming that a group of lambdas is applied
899 -- all at once. If there are too few args, we must zap the
900 -- occ-info, UNLESS the remaining binders are one-shot
901 where
902 (bndrs, body) = collectBinders expr
903 zapped_bndrs | need_to_zap = map zap bndrs
904 | otherwise = bndrs
905
906 need_to_zap = any zappable_bndr (drop n_args bndrs)
907 n_args = countArgs cont
908 -- NB: countArgs counts all the args (incl type args)
909 -- and likewise drop counts all binders (incl type lambdas)
910
911 zappable_bndr b = isId b && not (isOneShotBndr b)
912 zap b | isTyVar b = b
913 | otherwise = zapLamIdInfo b
914
915 simplExprF1 env (Case scrut bndr _ alts) cont
916 = simplExprF env scrut (Select { sc_dup = NoDup, sc_bndr = bndr
917 , sc_alts = alts
918 , sc_env = env, sc_cont = cont })
919
920 simplExprF1 env (Let (Rec pairs) body) cont
921 = do { env' <- simplRecBndrs env (map fst pairs)
922 -- NB: bndrs' don't have unfoldings or rules
923 -- We add them as we go down
924
925 ; env'' <- simplRecBind env' NotTopLevel pairs
926 ; simplExprF env'' body cont }
927
928 simplExprF1 env (Let (NonRec bndr rhs) body) cont
929 = simplNonRecE env bndr (rhs, env) ([], body) cont
930
931 ---------------------------------
932 simplType :: SimplEnv -> InType -> SimplM OutType
933 -- Kept monadic just so we can do the seqType
934 simplType env ty
935 = -- pprTrace "simplType" (ppr ty $$ ppr (seTvSubst env)) $
936 seqType new_ty `seq` return new_ty
937 where
938 new_ty = substTy env ty
939
940 ---------------------------------
941 simplCoercionF :: SimplEnv -> InCoercion -> SimplCont
942 -> SimplM (SimplEnv, OutExpr)
943 simplCoercionF env co cont
944 = do { co' <- simplCoercion env co
945 ; rebuild env (Coercion co') cont }
946
947 simplCoercion :: SimplEnv -> InCoercion -> SimplM OutCoercion
948 simplCoercion env co
949 = let opt_co = optCoercion (getTCvSubst env) co
950 in seqCo opt_co `seq` return opt_co
951
952 -----------------------------------
953 -- | Push a TickIt context outwards past applications and cases, as
954 -- long as this is a non-scoping tick, to let case and application
955 -- optimisations apply.
956
957 simplTick :: SimplEnv -> Tickish Id -> InExpr -> SimplCont
958 -> SimplM (SimplEnv, OutExpr)
959 simplTick env tickish expr cont
960 -- A scoped tick turns into a continuation, so that we can spot
961 -- (scc t (\x . e)) in simplLam and eliminate the scc. If we didn't do
962 -- it this way, then it would take two passes of the simplifier to
963 -- reduce ((scc t (\x . e)) e').
964 -- NB, don't do this with counting ticks, because if the expr is
965 -- bottom, then rebuildCall will discard the continuation.
966
967 -- XXX: we cannot do this, because the simplifier assumes that
968 -- the context can be pushed into a case with a single branch. e.g.
969 -- scc<f> case expensive of p -> e
970 -- becomes
971 -- case expensive of p -> scc<f> e
972 --
973 -- So I'm disabling this for now. It just means we will do more
974 -- simplifier iterations that necessary in some cases.
975
976 -- | tickishScoped tickish && not (tickishCounts tickish)
977 -- = simplExprF env expr (TickIt tickish cont)
978
979 -- For unscoped or soft-scoped ticks, we are allowed to float in new
980 -- cost, so we simply push the continuation inside the tick. This
981 -- has the effect of moving the tick to the outside of a case or
982 -- application context, allowing the normal case and application
983 -- optimisations to fire.
984 | tickish `tickishScopesLike` SoftScope
985 = do { (env', expr') <- simplExprF env expr cont
986 ; return (env', mkTick tickish expr')
987 }
988
989 -- Push tick inside if the context looks like this will allow us to
990 -- do a case-of-case - see Note [case-of-scc-of-case]
991 | Select {} <- cont, Just expr' <- push_tick_inside
992 = simplExprF env expr' cont
993
994 -- We don't want to move the tick, but we might still want to allow
995 -- floats to pass through with appropriate wrapping (or not, see
996 -- wrap_floats below)
997 --- | not (tickishCounts tickish) || tickishCanSplit tickish
998 -- = wrap_floats
999
1000 | otherwise
1001 = no_floating_past_tick
1002
1003 where
1004
1005 -- Try to push tick inside a case, see Note [case-of-scc-of-case].
1006 push_tick_inside =
1007 case expr0 of
1008 Case scrut bndr ty alts
1009 -> Just $ Case (tickScrut scrut) bndr ty (map tickAlt alts)
1010 _other -> Nothing
1011 where (ticks, expr0) = stripTicksTop movable (Tick tickish expr)
1012 movable t = not (tickishCounts t) ||
1013 t `tickishScopesLike` NoScope ||
1014 tickishCanSplit t
1015 tickScrut e = foldr mkTick e ticks
1016 -- Alternatives get annotated with all ticks that scope in some way,
1017 -- but we don't want to count entries.
1018 tickAlt (c,bs,e) = (c,bs, foldr mkTick e ts_scope)
1019 ts_scope = map mkNoCount $
1020 filter (not . (`tickishScopesLike` NoScope)) ticks
1021
1022 no_floating_past_tick =
1023 do { let (inc,outc) = splitCont cont
1024 ; (env', expr') <- simplExprF (zapFloats env) expr inc
1025 ; let tickish' = simplTickish env tickish
1026 ; (env'', expr'') <- rebuild (zapFloats env')
1027 (wrapFloats env' expr')
1028 (TickIt tickish' outc)
1029 ; return (addFloats env env'', expr'')
1030 }
1031
1032 -- Alternative version that wraps outgoing floats with the tick. This
1033 -- results in ticks being duplicated, as we don't make any attempt to
1034 -- eliminate the tick if we re-inline the binding (because the tick
1035 -- semantics allows unrestricted inlining of HNFs), so I'm not doing
1036 -- this any more. FloatOut will catch any real opportunities for
1037 -- floating.
1038 --
1039 -- wrap_floats =
1040 -- do { let (inc,outc) = splitCont cont
1041 -- ; (env', expr') <- simplExprF (zapFloats env) expr inc
1042 -- ; let tickish' = simplTickish env tickish
1043 -- ; let wrap_float (b,rhs) = (zapIdStrictness (setIdArity b 0),
1044 -- mkTick (mkNoCount tickish') rhs)
1045 -- -- when wrapping a float with mkTick, we better zap the Id's
1046 -- -- strictness info and arity, because it might be wrong now.
1047 -- ; let env'' = addFloats env (mapFloats env' wrap_float)
1048 -- ; rebuild env'' expr' (TickIt tickish' outc)
1049 -- }
1050
1051
1052 simplTickish env tickish
1053 | Breakpoint n ids <- tickish
1054 = Breakpoint n (map (getDoneId . substId env) ids)
1055 | otherwise = tickish
1056
1057 -- Push type application and coercion inside a tick
1058 splitCont :: SimplCont -> (SimplCont, SimplCont)
1059 splitCont cont@(ApplyToTy { sc_cont = tail }) = (cont { sc_cont = inc }, outc)
1060 where (inc,outc) = splitCont tail
1061 splitCont (CastIt co c) = (CastIt co inc, outc)
1062 where (inc,outc) = splitCont c
1063 splitCont other = (mkBoringStop (contHoleType other), other)
1064
1065 getDoneId (DoneId id) = id
1066 getDoneId (DoneEx e) = getIdFromTrivialExpr e -- Note [substTickish] in CoreSubst
1067 getDoneId other = pprPanic "getDoneId" (ppr other)
1068
1069 -- Note [case-of-scc-of-case]
1070 -- It's pretty important to be able to transform case-of-case when
1071 -- there's an SCC in the way. For example, the following comes up
1072 -- in nofib/real/compress/Encode.hs:
1073 --
1074 -- case scctick<code_string.r1>
1075 -- case $wcode_string_r13s wild_XC w1_s137 w2_s138 l_aje
1076 -- of _ { (# ww1_s13f, ww2_s13g, ww3_s13h #) ->
1077 -- (ww1_s13f, ww2_s13g, ww3_s13h)
1078 -- }
1079 -- of _ { (ww_s12Y, ww1_s12Z, ww2_s130) ->
1080 -- tick<code_string.f1>
1081 -- (ww_s12Y,
1082 -- ww1_s12Z,
1083 -- PTTrees.PT
1084 -- @ GHC.Types.Char @ GHC.Types.Int wild2_Xj ww2_s130 r_ajf)
1085 -- }
1086 --
1087 -- We really want this case-of-case to fire, because then the 3-tuple
1088 -- will go away (indeed, the CPR optimisation is relying on this
1089 -- happening). But the scctick is in the way - we need to push it
1090 -- inside to expose the case-of-case. So we perform this
1091 -- transformation on the inner case:
1092 --
1093 -- scctick c (case e of { p1 -> e1; ...; pn -> en })
1094 -- ==>
1095 -- case (scctick c e) of { p1 -> scc c e1; ...; pn -> scc c en }
1096 --
1097 -- So we've moved a constant amount of work out of the scc to expose
1098 -- the case. We only do this when the continuation is interesting: in
1099 -- for now, it has to be another Case (maybe generalise this later).
1100
1101 {-
1102 ************************************************************************
1103 * *
1104 \subsection{The main rebuilder}
1105 * *
1106 ************************************************************************
1107 -}
1108
1109 rebuild :: SimplEnv -> OutExpr -> SimplCont -> SimplM (SimplEnv, OutExpr)
1110 -- At this point the substitution in the SimplEnv should be irrelevant
1111 -- only the in-scope set and floats should matter
1112 rebuild env expr cont
1113 = case cont of
1114 Stop {} -> return (env, expr)
1115 TickIt t cont -> rebuild env (mkTick t expr) cont
1116 CastIt co cont -> rebuild env (mkCast expr co) cont
1117 -- NB: mkCast implements the (Coercion co |> g) optimisation
1118
1119 Select { sc_bndr = bndr, sc_alts = alts, sc_env = se, sc_cont = cont }
1120 -> rebuildCase (se `setFloats` env) expr bndr alts cont
1121
1122 StrictArg info _ cont -> rebuildCall env (info `addValArgTo` expr) cont
1123 StrictBind b bs body se cont -> do { env' <- simplNonRecX (se `setFloats` env) b expr
1124 -- expr satisfies let/app since it started life
1125 -- in a call to simplNonRecE
1126 ; simplLam env' bs body cont }
1127
1128 ApplyToTy { sc_arg_ty = ty, sc_cont = cont}
1129 -> rebuild env (App expr (Type ty)) cont
1130 ApplyToVal { sc_arg = arg, sc_env = se, sc_dup = dup_flag, sc_cont = cont}
1131 -- See Note [Avoid redundant simplification]
1132 | isSimplified dup_flag -> rebuild env (App expr arg) cont
1133 | otherwise -> do { arg' <- simplExpr (se `setInScope` env) arg
1134 ; rebuild env (App expr arg') cont }
1135
1136
1137 {-
1138 ************************************************************************
1139 * *
1140 \subsection{Lambdas}
1141 * *
1142 ************************************************************************
1143 -}
1144
1145 simplCast :: SimplEnv -> InExpr -> Coercion -> SimplCont
1146 -> SimplM (SimplEnv, OutExpr)
1147 simplCast env body co0 cont0
1148 = do { co1 <- simplCoercion env co0
1149 ; cont1 <- addCoerce co1 cont0
1150 ; simplExprF env body cont1 }
1151 where
1152 addCoerce :: OutCoercion -> SimplCont -> SimplM SimplCont
1153 addCoerce co1 (CastIt co2 cont)
1154 = addCoerce (mkTransCo co1 co2) cont
1155
1156 addCoerce co cont@(ApplyToTy { sc_arg_ty = arg_ty, sc_cont = tail })
1157 | Just (arg_ty', co') <- pushCoTyArg co arg_ty
1158 = do { tail' <- addCoerce co' tail
1159 ; return (cont { sc_arg_ty = arg_ty', sc_cont = tail' }) }
1160
1161 addCoerce co (ApplyToVal { sc_arg = arg, sc_env = arg_se
1162 , sc_dup = dup, sc_cont = tail })
1163 | Just (co1, co2) <- pushCoValArg co
1164 , Pair _ new_ty <- coercionKind co1
1165 , not (isTypeLevPoly new_ty) -- without this check, we get a lev-poly arg
1166 -- See Note [Levity polymorphism invariants] in CoreSyn
1167 -- test: typecheck/should_run/EtaExpandLevPoly
1168 = do { (dup', arg_se', arg') <- simplArg env dup arg_se arg
1169 -- When we build the ApplyTo we can't mix the OutCoercion
1170 -- 'co' with the InExpr 'arg', so we simplify
1171 -- to make it all consistent. It's a bit messy.
1172 -- But it isn't a common case.
1173 -- Example of use: Trac #995
1174 ; tail' <- addCoerce co2 tail
1175 ; return (ApplyToVal { sc_arg = mkCast arg' co1
1176 , sc_env = arg_se'
1177 , sc_dup = dup'
1178 , sc_cont = tail' }) }
1179
1180 addCoerce co cont
1181 | isReflexiveCo co = return cont
1182 | otherwise = return (CastIt co cont)
1183 -- It's worth checking isReflexiveCo.
1184 -- For example, in the initial form of a worker
1185 -- we may find (coerce T (coerce S (\x.e))) y
1186 -- and we'd like it to simplify to e[y/x] in one round
1187 -- of simplification
1188
1189 simplArg :: SimplEnv -> DupFlag -> StaticEnv -> CoreExpr
1190 -> SimplM (DupFlag, StaticEnv, OutExpr)
1191 simplArg env dup_flag arg_env arg
1192 | isSimplified dup_flag
1193 = return (dup_flag, arg_env, arg)
1194 | otherwise
1195 = do { arg' <- simplExpr (arg_env `setInScope` env) arg
1196 ; return (Simplified, zapSubstEnv arg_env, arg') }
1197
1198 {-
1199 ************************************************************************
1200 * *
1201 \subsection{Lambdas}
1202 * *
1203 ************************************************************************
1204
1205 Note [Zap unfolding when beta-reducing]
1206 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1207 Lambda-bound variables can have stable unfoldings, such as
1208 $j = \x. \b{Unf=Just x}. e
1209 See Note [Case binders and join points] below; the unfolding for lets
1210 us optimise e better. However when we beta-reduce it we want to
1211 revert to using the actual value, otherwise we can end up in the
1212 stupid situation of
1213 let x = blah in
1214 let b{Unf=Just x} = y
1215 in ...b...
1216 Here it'd be far better to drop the unfolding and use the actual RHS.
1217 -}
1218
1219 simplLam :: SimplEnv -> [InId] -> InExpr -> SimplCont
1220 -> SimplM (SimplEnv, OutExpr)
1221
1222 simplLam env [] body cont = simplExprF env body cont
1223
1224 -- Beta reduction
1225
1226 simplLam env (bndr:bndrs) body (ApplyToTy { sc_arg_ty = arg_ty, sc_cont = cont })
1227 = do { tick (BetaReduction bndr)
1228 ; simplLam (extendTvSubst env bndr arg_ty) bndrs body cont }
1229
1230 simplLam env (bndr:bndrs) body (ApplyToVal { sc_arg = arg, sc_env = arg_se
1231 , sc_cont = cont })
1232 = do { tick (BetaReduction bndr)
1233 ; simplNonRecE env' (zap_unfolding bndr) (arg, arg_se) (bndrs, body) cont }
1234 where
1235 env' | Coercion co <- arg
1236 = extendCvSubst env bndr co
1237 | otherwise
1238 = env
1239
1240 zap_unfolding bndr -- See Note [Zap unfolding when beta-reducing]
1241 | isId bndr, isStableUnfolding (realIdUnfolding bndr)
1242 = setIdUnfolding bndr NoUnfolding
1243 | otherwise = bndr
1244
1245 -- discard a non-counting tick on a lambda. This may change the
1246 -- cost attribution slightly (moving the allocation of the
1247 -- lambda elsewhere), but we don't care: optimisation changes
1248 -- cost attribution all the time.
1249 simplLam env bndrs body (TickIt tickish cont)
1250 | not (tickishCounts tickish)
1251 = simplLam env bndrs body cont
1252
1253 -- Not enough args, so there are real lambdas left to put in the result
1254 simplLam env bndrs body cont
1255 = do { (env', bndrs') <- simplLamBndrs env bndrs
1256 ; body' <- simplExpr env' body
1257 ; new_lam <- mkLam env bndrs' body' cont
1258 ; rebuild env' new_lam cont }
1259
1260 simplLamBndrs :: SimplEnv -> [InBndr] -> SimplM (SimplEnv, [OutBndr])
1261 simplLamBndrs env bndrs = mapAccumLM simplLamBndr env bndrs
1262
1263 -------------
1264 simplLamBndr :: SimplEnv -> Var -> SimplM (SimplEnv, Var)
1265 -- Used for lambda binders. These sometimes have unfoldings added by
1266 -- the worker/wrapper pass that must be preserved, because they can't
1267 -- be reconstructed from context. For example:
1268 -- f x = case x of (a,b) -> fw a b x
1269 -- fw a b x{=(a,b)} = ...
1270 -- The "{=(a,b)}" is an unfolding we can't reconstruct otherwise.
1271 simplLamBndr env bndr
1272 | isId bndr && hasSomeUnfolding old_unf -- Special case
1273 = do { (env1, bndr1) <- simplBinder env bndr
1274 ; unf' <- simplUnfolding env1 NotTopLevel bndr old_unf
1275 ; let bndr2 = bndr1 `setIdUnfolding` unf'
1276 ; return (modifyInScope env1 bndr2, bndr2) }
1277
1278 | otherwise
1279 = simplBinder env bndr -- Normal case
1280 where
1281 old_unf = idUnfolding bndr
1282
1283 ------------------
1284 simplNonRecE :: SimplEnv
1285 -> InBndr -- The binder
1286 -> (InExpr, SimplEnv) -- Rhs of binding (or arg of lambda)
1287 -> ([InBndr], InExpr) -- Body of the let/lambda
1288 -- \xs.e
1289 -> SimplCont
1290 -> SimplM (SimplEnv, OutExpr)
1291
1292 -- simplNonRecE is used for
1293 -- * non-top-level non-recursive lets in expressions
1294 -- * beta reduction
1295 --
1296 -- It deals with strict bindings, via the StrictBind continuation,
1297 -- which may abort the whole process
1298 --
1299 -- Precondition: rhs satisfies the let/app invariant
1300 -- Note [CoreSyn let/app invariant] in CoreSyn
1301 --
1302 -- The "body" of the binding comes as a pair of ([InId],InExpr)
1303 -- representing a lambda; so we recurse back to simplLam
1304 -- Why? Because of the binder-occ-info-zapping done before
1305 -- the call to simplLam in simplExprF (Lam ...)
1306
1307 -- First deal with type applications and type lets
1308 -- (/\a. e) (Type ty) and (let a = Type ty in e)
1309 simplNonRecE env bndr (Type ty_arg, rhs_se) (bndrs, body) cont
1310 = ASSERT( isTyVar bndr )
1311 do { ty_arg' <- simplType (rhs_se `setInScope` env) ty_arg
1312 ; simplLam (extendTvSubst env bndr ty_arg') bndrs body cont }
1313
1314 simplNonRecE env bndr (rhs, rhs_se) (bndrs, body) cont
1315 = do dflags <- getDynFlags
1316 case () of
1317 _ | preInlineUnconditionally dflags env NotTopLevel bndr rhs
1318 -> do { tick (PreInlineUnconditionally bndr)
1319 ; -- pprTrace "preInlineUncond" (ppr bndr <+> ppr rhs) $
1320 simplLam (extendIdSubst env bndr (mkContEx rhs_se rhs)) bndrs body cont }
1321
1322 | isStrictId bndr -- Includes coercions
1323 -> simplExprF (rhs_se `setFloats` env) rhs
1324 (StrictBind bndr bndrs body env cont)
1325
1326 | otherwise
1327 -> ASSERT( not (isTyVar bndr) )
1328 do { (env1, bndr1) <- simplNonRecBndr env bndr
1329 ; (env2, bndr2) <- addBndrRules env1 bndr bndr1
1330 ; env3 <- simplLazyBind env2 NotTopLevel NonRecursive bndr bndr2 rhs rhs_se
1331 ; simplLam env3 bndrs body cont }
1332
1333 {-
1334 ************************************************************************
1335 * *
1336 Variables
1337 * *
1338 ************************************************************************
1339 -}
1340
1341 simplVar :: SimplEnv -> InVar -> SimplM OutExpr
1342 -- Look up an InVar in the environment
1343 simplVar env var
1344 | isTyVar var = return (Type (substTyVar env var))
1345 | isCoVar var = return (Coercion (substCoVar env var))
1346 | otherwise
1347 = case substId env var of
1348 DoneId var1 -> return (Var var1)
1349 DoneEx e -> return e
1350 ContEx tvs cvs ids e -> simplExpr (setSubstEnv env tvs cvs ids) e
1351
1352 simplIdF :: SimplEnv -> InId -> SimplCont -> SimplM (SimplEnv, OutExpr)
1353 simplIdF env var cont
1354 = case substId env var of
1355 DoneEx e -> simplExprF (zapSubstEnv env) e cont
1356 ContEx tvs cvs ids e -> simplExprF (setSubstEnv env tvs cvs ids) e cont
1357 DoneId var1 -> completeCall env var1 cont
1358 -- Note [zapSubstEnv]
1359 -- The template is already simplified, so don't re-substitute.
1360 -- This is VITAL. Consider
1361 -- let x = e in
1362 -- let y = \z -> ...x... in
1363 -- \ x -> ...y...
1364 -- We'll clone the inner \x, adding x->x' in the id_subst
1365 -- Then when we inline y, we must *not* replace x by x' in
1366 -- the inlined copy!!
1367
1368 ---------------------------------------------------------
1369 -- Dealing with a call site
1370
1371 completeCall :: SimplEnv -> OutId -> SimplCont -> SimplM (SimplEnv, OutExpr)
1372 completeCall env var cont
1373 = do { ------------- Try inlining ----------------
1374 dflags <- getDynFlags
1375 ; let (lone_variable, arg_infos, call_cont) = contArgs cont
1376 n_val_args = length arg_infos
1377 interesting_cont = interestingCallContext call_cont
1378 unfolding = activeUnfolding env var
1379 maybe_inline = callSiteInline dflags var unfolding
1380 lone_variable arg_infos interesting_cont
1381 ; case maybe_inline of {
1382 Just expr -- There is an inlining!
1383 -> do { checkedTick (UnfoldingDone var)
1384 ; dump_inline dflags expr cont
1385 ; simplExprF (zapSubstEnv env) expr cont }
1386
1387 ; Nothing -> do -- No inlining!
1388
1389 { rule_base <- getSimplRules
1390 ; let info = mkArgInfo var (getRules rule_base var) n_val_args call_cont
1391 ; rebuildCall env info cont
1392 }}}
1393 where
1394 dump_inline dflags unfolding cont
1395 | not (dopt Opt_D_dump_inlinings dflags) = return ()
1396 | not (dopt Opt_D_verbose_core2core dflags)
1397 = when (isExternalName (idName var)) $
1398 liftIO $ printOutputForUser dflags alwaysQualify $
1399 sep [text "Inlining done:", nest 4 (ppr var)]
1400 | otherwise
1401 = liftIO $ printOutputForUser dflags alwaysQualify $
1402 sep [text "Inlining done: " <> ppr var,
1403 nest 4 (vcat [text "Inlined fn: " <+> nest 2 (ppr unfolding),
1404 text "Cont: " <+> ppr cont])]
1405
1406 rebuildCall :: SimplEnv
1407 -> ArgInfo
1408 -> SimplCont
1409 -> SimplM (SimplEnv, OutExpr)
1410 rebuildCall env (ArgInfo { ai_fun = fun, ai_args = rev_args, ai_strs = [] }) cont
1411 -- When we run out of strictness args, it means
1412 -- that the call is definitely bottom; see SimplUtils.mkArgInfo
1413 -- Then we want to discard the entire strict continuation. E.g.
1414 -- * case (error "hello") of { ... }
1415 -- * (error "Hello") arg
1416 -- * f (error "Hello") where f is strict
1417 -- etc
1418 -- Then, especially in the first of these cases, we'd like to discard
1419 -- the continuation, leaving just the bottoming expression. But the
1420 -- type might not be right, so we may have to add a coerce.
1421 | not (contIsTrivial cont) -- Only do this if there is a non-trivial
1422 = return (env, castBottomExpr res cont_ty) -- continuation to discard, else we do it
1423 where -- again and again!
1424 res = argInfoExpr fun rev_args
1425 cont_ty = contResultType cont
1426
1427 rebuildCall env info (CastIt co cont)
1428 = rebuildCall env (addCastTo info co) cont
1429
1430 rebuildCall env info (ApplyToTy { sc_arg_ty = arg_ty, sc_cont = cont })
1431 = rebuildCall env (info `addTyArgTo` arg_ty) cont
1432
1433 rebuildCall env info@(ArgInfo { ai_encl = encl_rules, ai_type = fun_ty
1434 , ai_strs = str:strs, ai_discs = disc:discs })
1435 (ApplyToVal { sc_arg = arg, sc_env = arg_se
1436 , sc_dup = dup_flag, sc_cont = cont })
1437 | isSimplified dup_flag -- See Note [Avoid redundant simplification]
1438 = rebuildCall env (addValArgTo info' arg) cont
1439
1440 | str -- Strict argument
1441 = -- pprTrace "Strict Arg" (ppr arg $$ ppr (seIdSubst env) $$ ppr (seInScope env)) $
1442 simplExprF (arg_se `setFloats` env) arg
1443 (StrictArg info' cci cont)
1444 -- Note [Shadowing]
1445
1446 | otherwise -- Lazy argument
1447 -- DO NOT float anything outside, hence simplExprC
1448 -- There is no benefit (unlike in a let-binding), and we'd
1449 -- have to be very careful about bogus strictness through
1450 -- floating a demanded let.
1451 = do { arg' <- simplExprC (arg_se `setInScope` env) arg
1452 (mkLazyArgStop (funArgTy fun_ty) cci)
1453 ; rebuildCall env (addValArgTo info' arg') cont }
1454 where
1455 info' = info { ai_strs = strs, ai_discs = discs }
1456 cci | encl_rules = RuleArgCtxt
1457 | disc > 0 = DiscArgCtxt -- Be keener here
1458 | otherwise = BoringCtxt -- Nothing interesting
1459
1460 rebuildCall env (ArgInfo { ai_fun = fun, ai_args = rev_args, ai_rules = rules }) cont
1461 | null rules
1462 = rebuild env (argInfoExpr fun rev_args) cont -- No rules, common case
1463
1464 | otherwise
1465 = do { -- We've accumulated a simplified call in <fun,rev_args>
1466 -- so try rewrite rules; see Note [RULEs apply to simplified arguments]
1467 -- See also Note [Rules for recursive functions]
1468 ; let env' = zapSubstEnv env -- See Note [zapSubstEnv];
1469 -- and NB that 'rev_args' are all fully simplified
1470 ; mb_rule <- tryRules env' rules fun (reverse rev_args) cont
1471 ; case mb_rule of {
1472 Just (rule_rhs, cont') -> simplExprF env' rule_rhs cont'
1473
1474 -- Rules don't match
1475 ; Nothing -> rebuild env (argInfoExpr fun rev_args) cont -- No rules
1476 } }
1477
1478 {-
1479 Note [RULES apply to simplified arguments]
1480 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1481 It's very desirable to try RULES once the arguments have been simplified, because
1482 doing so ensures that rule cascades work in one pass. Consider
1483 {-# RULES g (h x) = k x
1484 f (k x) = x #-}
1485 ...f (g (h x))...
1486 Then we want to rewrite (g (h x)) to (k x) and only then try f's rules. If
1487 we match f's rules against the un-simplified RHS, it won't match. This
1488 makes a particularly big difference when superclass selectors are involved:
1489 op ($p1 ($p2 (df d)))
1490 We want all this to unravel in one sweep.
1491
1492 Note [Avoid redundant simplification]
1493 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1494 Because RULES apply to simplified arguments, there's a danger of repeatedly
1495 simplifying already-simplified arguments. An important example is that of
1496 (>>=) d e1 e2
1497 Here e1, e2 are simplified before the rule is applied, but don't really
1498 participate in the rule firing. So we mark them as Simplified to avoid
1499 re-simplifying them.
1500
1501 Note [Shadowing]
1502 ~~~~~~~~~~~~~~~~
1503 This part of the simplifier may break the no-shadowing invariant
1504 Consider
1505 f (...(\a -> e)...) (case y of (a,b) -> e')
1506 where f is strict in its second arg
1507 If we simplify the innermost one first we get (...(\a -> e)...)
1508 Simplifying the second arg makes us float the case out, so we end up with
1509 case y of (a,b) -> f (...(\a -> e)...) e'
1510 So the output does not have the no-shadowing invariant. However, there is
1511 no danger of getting name-capture, because when the first arg was simplified
1512 we used an in-scope set that at least mentioned all the variables free in its
1513 static environment, and that is enough.
1514
1515 We can't just do innermost first, or we'd end up with a dual problem:
1516 case x of (a,b) -> f e (...(\a -> e')...)
1517
1518 I spent hours trying to recover the no-shadowing invariant, but I just could
1519 not think of an elegant way to do it. The simplifier is already knee-deep in
1520 continuations. We have to keep the right in-scope set around; AND we have
1521 to get the effect that finding (error "foo") in a strict arg position will
1522 discard the entire application and replace it with (error "foo"). Getting
1523 all this at once is TOO HARD!
1524
1525
1526 ************************************************************************
1527 * *
1528 Rewrite rules
1529 * *
1530 ************************************************************************
1531 -}
1532
1533 tryRules :: SimplEnv -> [CoreRule]
1534 -> Id -> [ArgSpec] -> SimplCont
1535 -> SimplM (Maybe (CoreExpr, SimplCont))
1536 -- The SimplEnv already has zapSubstEnv applied to it
1537
1538 tryRules env rules fn args call_cont
1539 | null rules
1540 = return Nothing
1541 {- Disabled until we fix #8326
1542 | fn `hasKey` tagToEnumKey -- See Note [Optimising tagToEnum#]
1543 , [_type_arg, val_arg] <- args
1544 , Select dup bndr ((_,[],rhs1) : rest_alts) se cont <- call_cont
1545 , isDeadBinder bndr
1546 = do { dflags <- getDynFlags
1547 ; let enum_to_tag :: CoreAlt -> CoreAlt
1548 -- Takes K -> e into tagK# -> e
1549 -- where tagK# is the tag of constructor K
1550 enum_to_tag (DataAlt con, [], rhs)
1551 = ASSERT( isEnumerationTyCon (dataConTyCon con) )
1552 (LitAlt tag, [], rhs)
1553 where
1554 tag = mkMachInt dflags (toInteger (dataConTag con - fIRST_TAG))
1555 enum_to_tag alt = pprPanic "tryRules: tagToEnum" (ppr alt)
1556
1557 new_alts = (DEFAULT, [], rhs1) : map enum_to_tag rest_alts
1558 new_bndr = setIdType bndr intPrimTy
1559 -- The binder is dead, but should have the right type
1560 ; return (Just (val_arg, Select dup new_bndr new_alts se cont)) }
1561 -}
1562 | otherwise
1563 = do { dflags <- getDynFlags
1564 ; case lookupRule dflags (getUnfoldingInRuleMatch env) (activeRule env)
1565 fn (argInfoAppArgs args) rules of {
1566 Nothing ->
1567 do { nodump dflags -- This ensures that an empty file is written
1568 ; return Nothing } ; -- No rule matches
1569 Just (rule, rule_rhs) ->
1570 do { checkedTick (RuleFired (ru_name rule))
1571 ; let cont' = pushSimplifiedArgs env
1572 (drop (ruleArity rule) args)
1573 call_cont
1574 -- (ruleArity rule) says how many args the rule consumed
1575 ; dump dflags rule rule_rhs
1576 ; return (Just (rule_rhs, cont')) }}}
1577 where
1578 dump dflags rule rule_rhs
1579 | dopt Opt_D_dump_rule_rewrites dflags
1580 = log_rule dflags Opt_D_dump_rule_rewrites "Rule fired" $ vcat
1581 [ text "Rule:" <+> ftext (ru_name rule)
1582 , text "Before:" <+> hang (ppr fn) 2 (sep (map ppr args))
1583 , text "After: " <+> pprCoreExpr rule_rhs
1584 , text "Cont: " <+> ppr call_cont ]
1585
1586 | dopt Opt_D_dump_rule_firings dflags
1587 = log_rule dflags Opt_D_dump_rule_firings "Rule fired:" $
1588 ftext (ru_name rule)
1589
1590 | otherwise
1591 = return ()
1592
1593 nodump dflags
1594 | dopt Opt_D_dump_rule_rewrites dflags
1595 = liftIO $ dumpSDoc dflags alwaysQualify Opt_D_dump_rule_rewrites "" empty
1596
1597 | dopt Opt_D_dump_rule_firings dflags
1598 = liftIO $ dumpSDoc dflags alwaysQualify Opt_D_dump_rule_firings "" empty
1599
1600 | otherwise
1601 = return ()
1602
1603 log_rule dflags flag hdr details
1604 = liftIO . dumpSDoc dflags alwaysQualify flag "" $
1605 sep [text hdr, nest 4 details]
1606
1607 {-
1608 Note [Optimising tagToEnum#]
1609 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1610 If we have an enumeration data type:
1611
1612 data Foo = A | B | C
1613
1614 Then we want to transform
1615
1616 case tagToEnum# x of ==> case x of
1617 A -> e1 DEFAULT -> e1
1618 B -> e2 1# -> e2
1619 C -> e3 2# -> e3
1620
1621 thereby getting rid of the tagToEnum# altogether. If there was a DEFAULT
1622 alternative we retain it (remember it comes first). If not the case must
1623 be exhaustive, and we reflect that in the transformed version by adding
1624 a DEFAULT. Otherwise Lint complains that the new case is not exhaustive.
1625 See #8317.
1626
1627 Note [Rules for recursive functions]
1628 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1629 You might think that we shouldn't apply rules for a loop breaker:
1630 doing so might give rise to an infinite loop, because a RULE is
1631 rather like an extra equation for the function:
1632 RULE: f (g x) y = x+y
1633 Eqn: f a y = a-y
1634
1635 But it's too drastic to disable rules for loop breakers.
1636 Even the foldr/build rule would be disabled, because foldr
1637 is recursive, and hence a loop breaker:
1638 foldr k z (build g) = g k z
1639 So it's up to the programmer: rules can cause divergence
1640
1641
1642 ************************************************************************
1643 * *
1644 Rebuilding a case expression
1645 * *
1646 ************************************************************************
1647
1648 Note [Case elimination]
1649 ~~~~~~~~~~~~~~~~~~~~~~~
1650 The case-elimination transformation discards redundant case expressions.
1651 Start with a simple situation:
1652
1653 case x# of ===> let y# = x# in e
1654 y# -> e
1655
1656 (when x#, y# are of primitive type, of course). We can't (in general)
1657 do this for algebraic cases, because we might turn bottom into
1658 non-bottom!
1659
1660 The code in SimplUtils.prepareAlts has the effect of generalise this
1661 idea to look for a case where we're scrutinising a variable, and we
1662 know that only the default case can match. For example:
1663
1664 case x of
1665 0# -> ...
1666 DEFAULT -> ...(case x of
1667 0# -> ...
1668 DEFAULT -> ...) ...
1669
1670 Here the inner case is first trimmed to have only one alternative, the
1671 DEFAULT, after which it's an instance of the previous case. This
1672 really only shows up in eliminating error-checking code.
1673
1674 Note that SimplUtils.mkCase combines identical RHSs. So
1675
1676 case e of ===> case e of DEFAULT -> r
1677 True -> r
1678 False -> r
1679
1680 Now again the case may be elminated by the CaseElim transformation.
1681 This includes things like (==# a# b#)::Bool so that we simplify
1682 case ==# a# b# of { True -> x; False -> x }
1683 to just
1684 x
1685 This particular example shows up in default methods for
1686 comparison operations (e.g. in (>=) for Int.Int32)
1687
1688 Note [Case elimination: lifted case]
1689 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1690 If a case over a lifted type has a single alternative, and is being used
1691 as a strict 'let' (all isDeadBinder bndrs), we may want to do this
1692 transformation:
1693
1694 case e of r ===> let r = e in ...r...
1695 _ -> ...r...
1696
1697 (a) 'e' is already evaluated (it may so if e is a variable)
1698 Specifically we check (exprIsHNF e). In this case
1699 we can just allocate the WHNF directly with a let.
1700 or
1701 (b) 'x' is not used at all and e is ok-for-speculation
1702 The ok-for-spec bit checks that we don't lose any
1703 exceptions or divergence.
1704
1705 NB: it'd be *sound* to switch from case to let if the
1706 scrutinee was not yet WHNF but was guaranteed to
1707 converge; but sticking with case means we won't build a
1708 thunk
1709
1710 or
1711 (c) 'x' is used strictly in the body, and 'e' is a variable
1712 Then we can just substitute 'e' for 'x' in the body.
1713 See Note [Eliminating redundant seqs]
1714
1715 For (b), the "not used at all" test is important. Consider
1716 case (case a ># b of { True -> (p,q); False -> (q,p) }) of
1717 r -> blah
1718 The scrutinee is ok-for-speculation (it looks inside cases), but we do
1719 not want to transform to
1720 let r = case a ># b of { True -> (p,q); False -> (q,p) }
1721 in blah
1722 because that builds an unnecessary thunk.
1723
1724 Note [Eliminating redundant seqs]
1725 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1726 If we have this:
1727 case x of r { _ -> ..r.. }
1728 where 'r' is used strictly in (..r..), the case is effectively a 'seq'
1729 on 'x', but since 'r' is used strictly anyway, we can safely transform to
1730 (...x...)
1731
1732 Note that this can change the error behaviour. For example, we might
1733 transform
1734 case x of { _ -> error "bad" }
1735 --> error "bad"
1736 which is might be puzzling if 'x' currently lambda-bound, but later gets
1737 let-bound to (error "good").
1738
1739 Nevertheless, the paper "A semantics for imprecise exceptions" allows
1740 this transformation. If you want to fix the evaluation order, use
1741 'pseq'. See Trac #8900 for an example where the loss of this
1742 transformation bit us in practice.
1743
1744 See also Note [Empty case alternatives] in CoreSyn.
1745
1746 Just for reference, the original code (added Jan 13) looked like this:
1747 || case_bndr_evald_next rhs
1748
1749 case_bndr_evald_next :: CoreExpr -> Bool
1750 -- See Note [Case binder next]
1751 case_bndr_evald_next (Var v) = v == case_bndr
1752 case_bndr_evald_next (Cast e _) = case_bndr_evald_next e
1753 case_bndr_evald_next (App e _) = case_bndr_evald_next e
1754 case_bndr_evald_next (Case e _ _ _) = case_bndr_evald_next e
1755 case_bndr_evald_next _ = False
1756
1757 (This came up when fixing Trac #7542. See also Note [Eta reduction of
1758 an eval'd function] in CoreUtils.)
1759
1760
1761 Note [Case elimination: unlifted case]
1762 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1763 Consider
1764 case a +# b of r -> ...r...
1765 Then we do case-elimination (to make a let) followed by inlining,
1766 to get
1767 .....(a +# b)....
1768 If we have
1769 case indexArray# a i of r -> ...r...
1770 we might like to do the same, and inline the (indexArray# a i).
1771 But indexArray# is not okForSpeculation, so we don't build a let
1772 in rebuildCase (lest it get floated *out*), so the inlining doesn't
1773 happen either.
1774
1775 This really isn't a big deal I think. The let can be
1776
1777
1778 Further notes about case elimination
1779 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1780 Consider: test :: Integer -> IO ()
1781 test = print
1782
1783 Turns out that this compiles to:
1784 Print.test
1785 = \ eta :: Integer
1786 eta1 :: Void# ->
1787 case PrelNum.< eta PrelNum.zeroInteger of wild { __DEFAULT ->
1788 case hPutStr stdout
1789 (PrelNum.jtos eta ($w[] @ Char))
1790 eta1
1791 of wild1 { (# new_s, a4 #) -> PrelIO.lvl23 new_s }}
1792
1793 Notice the strange '<' which has no effect at all. This is a funny one.
1794 It started like this:
1795
1796 f x y = if x < 0 then jtos x
1797 else if y==0 then "" else jtos x
1798
1799 At a particular call site we have (f v 1). So we inline to get
1800
1801 if v < 0 then jtos x
1802 else if 1==0 then "" else jtos x
1803
1804 Now simplify the 1==0 conditional:
1805
1806 if v<0 then jtos v else jtos v
1807
1808 Now common-up the two branches of the case:
1809
1810 case (v<0) of DEFAULT -> jtos v
1811
1812 Why don't we drop the case? Because it's strict in v. It's technically
1813 wrong to drop even unnecessary evaluations, and in practice they
1814 may be a result of 'seq' so we *definitely* don't want to drop those.
1815 I don't really know how to improve this situation.
1816 -}
1817
1818 ---------------------------------------------------------
1819 -- Eliminate the case if possible
1820
1821 rebuildCase, reallyRebuildCase
1822 :: SimplEnv
1823 -> OutExpr -- Scrutinee
1824 -> InId -- Case binder
1825 -> [InAlt] -- Alternatives (inceasing order)
1826 -> SimplCont
1827 -> SimplM (SimplEnv, OutExpr)
1828
1829 --------------------------------------------------
1830 -- 1. Eliminate the case if there's a known constructor
1831 --------------------------------------------------
1832
1833 rebuildCase env scrut case_bndr alts cont
1834 | Lit lit <- scrut -- No need for same treatment as constructors
1835 -- because literals are inlined more vigorously
1836 , not (litIsLifted lit)
1837 = do { tick (KnownBranch case_bndr)
1838 ; case findAlt (LitAlt lit) alts of
1839 Nothing -> missingAlt env case_bndr alts cont
1840 Just (_, bs, rhs) -> simple_rhs bs rhs }
1841
1842 | Just (con, ty_args, other_args) <- exprIsConApp_maybe (getUnfoldingInRuleMatch env) scrut
1843 -- Works when the scrutinee is a variable with a known unfolding
1844 -- as well as when it's an explicit constructor application
1845 = do { tick (KnownBranch case_bndr)
1846 ; case findAlt (DataAlt con) alts of
1847 Nothing -> missingAlt env case_bndr alts cont
1848 Just (DEFAULT, bs, rhs) -> simple_rhs bs rhs
1849 Just (_, bs, rhs) -> knownCon env scrut con ty_args other_args
1850 case_bndr bs rhs cont
1851 }
1852 where
1853 simple_rhs bs rhs = ASSERT( null bs )
1854 do { env' <- simplNonRecX env case_bndr scrut
1855 -- scrut is a constructor application,
1856 -- hence satisfies let/app invariant
1857 ; simplExprF env' rhs cont }
1858
1859
1860 --------------------------------------------------
1861 -- 2. Eliminate the case if scrutinee is evaluated
1862 --------------------------------------------------
1863
1864 rebuildCase env scrut case_bndr alts@[(_, bndrs, rhs)] cont
1865 -- See if we can get rid of the case altogether
1866 -- See Note [Case elimination]
1867 -- mkCase made sure that if all the alternatives are equal,
1868 -- then there is now only one (DEFAULT) rhs
1869
1870 -- 2a. Dropping the case altogether, if
1871 -- a) it binds nothing (so it's really just a 'seq')
1872 -- b) evaluating the scrutinee has no side effects
1873 | is_plain_seq
1874 , exprOkForSideEffects scrut
1875 -- The entire case is dead, so we can drop it
1876 -- if the scrutinee converges without having imperative
1877 -- side effects or raising a Haskell exception
1878 -- See Note [PrimOp can_fail and has_side_effects] in PrimOp
1879 = simplExprF env rhs cont
1880
1881 -- 2b. Turn the case into a let, if
1882 -- a) it binds only the case-binder
1883 -- b) unlifted case: the scrutinee is ok-for-speculation
1884 -- lifted case: the scrutinee is in HNF (or will later be demanded)
1885 | all_dead_bndrs
1886 , if is_unlifted
1887 then exprOkForSpeculation scrut -- See Note [Case elimination: unlifted case]
1888 else exprIsHNF scrut -- See Note [Case elimination: lifted case]
1889 || scrut_is_demanded_var scrut
1890 = do { tick (CaseElim case_bndr)
1891 ; env' <- simplNonRecX env case_bndr scrut
1892 ; simplExprF env' rhs cont }
1893
1894 -- 2c. Try the seq rules if
1895 -- a) it binds only the case binder
1896 -- b) a rule for seq applies
1897 -- See Note [User-defined RULES for seq] in MkId
1898 | is_plain_seq
1899 = do { let scrut_ty = exprType scrut
1900 rhs_ty = substTy env (exprType rhs)
1901 out_args = [ TyArg { as_arg_ty = scrut_ty
1902 , as_hole_ty = seq_id_ty }
1903 , TyArg { as_arg_ty = rhs_ty
1904 , as_hole_ty = piResultTy seq_id_ty scrut_ty }
1905 , ValArg scrut]
1906 rule_cont = ApplyToVal { sc_dup = NoDup, sc_arg = rhs
1907 , sc_env = env, sc_cont = cont }
1908 env' = zapSubstEnv env
1909 -- Lazily evaluated, so we don't do most of this
1910
1911 ; rule_base <- getSimplRules
1912 ; mb_rule <- tryRules env' (getRules rule_base seqId) seqId out_args rule_cont
1913 ; case mb_rule of
1914 Just (rule_rhs, cont') -> simplExprF env' rule_rhs cont'
1915 Nothing -> reallyRebuildCase env scrut case_bndr alts cont }
1916 where
1917 is_unlifted = isUnliftedType (idType case_bndr)
1918 all_dead_bndrs = all isDeadBinder bndrs -- bndrs are [InId]
1919 is_plain_seq = all_dead_bndrs && isDeadBinder case_bndr -- Evaluation *only* for effect
1920 seq_id_ty = idType seqId
1921
1922 scrut_is_demanded_var :: CoreExpr -> Bool
1923 -- See Note [Eliminating redundant seqs]
1924 scrut_is_demanded_var (Cast s _) = scrut_is_demanded_var s
1925 scrut_is_demanded_var (Var _) = isStrictDmd (idDemandInfo case_bndr)
1926 scrut_is_demanded_var _ = False
1927
1928
1929 rebuildCase env scrut case_bndr alts cont
1930 = reallyRebuildCase env scrut case_bndr alts cont
1931
1932 --------------------------------------------------
1933 -- 3. Catch-all case
1934 --------------------------------------------------
1935
1936 reallyRebuildCase env scrut case_bndr alts cont
1937 = do { -- Prepare the continuation;
1938 -- The new subst_env is in place
1939 (env', dup_cont, nodup_cont) <- prepareCaseCont env alts cont
1940
1941 -- Simplify the alternatives
1942 ; (scrut', case_bndr', alts') <- simplAlts env' scrut case_bndr alts dup_cont
1943
1944 ; dflags <- getDynFlags
1945 ; let alts_ty' = contResultType dup_cont
1946 ; case_expr <- mkCase dflags scrut' case_bndr' alts_ty' alts'
1947
1948 -- Notice that rebuild gets the in-scope set from env', not alt_env
1949 -- (which in any case is only build in simplAlts)
1950 -- The case binder *not* scope over the whole returned case-expression
1951 ; rebuild env' case_expr nodup_cont }
1952
1953 {-
1954 simplCaseBinder checks whether the scrutinee is a variable, v. If so,
1955 try to eliminate uses of v in the RHSs in favour of case_bndr; that
1956 way, there's a chance that v will now only be used once, and hence
1957 inlined.
1958
1959 Historical note: we use to do the "case binder swap" in the Simplifier
1960 so there were additional complications if the scrutinee was a variable.
1961 Now the binder-swap stuff is done in the occurrence analyer; see
1962 OccurAnal Note [Binder swap].
1963
1964 Note [knownCon occ info]
1965 ~~~~~~~~~~~~~~~~~~~~~~~~
1966 If the case binder is not dead, then neither are the pattern bound
1967 variables:
1968 case <any> of x { (a,b) ->
1969 case x of { (p,q) -> p } }
1970 Here (a,b) both look dead, but come alive after the inner case is eliminated.
1971 The point is that we bring into the envt a binding
1972 let x = (a,b)
1973 after the outer case, and that makes (a,b) alive. At least we do unless
1974 the case binder is guaranteed dead.
1975
1976 Note [Case alternative occ info]
1977 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1978 When we are simply reconstructing a case (the common case), we always
1979 zap the occurrence info on the binders in the alternatives. Even
1980 if the case binder is dead, the scrutinee is usually a variable, and *that*
1981 can bring the case-alternative binders back to life.
1982 See Note [Add unfolding for scrutinee]
1983
1984 Note [Improving seq]
1985 ~~~~~~~~~~~~~~~~~~~
1986 Consider
1987 type family F :: * -> *
1988 type instance F Int = Int
1989
1990 ... case e of x { DEFAULT -> rhs } ...
1991
1992 where x::F Int. Then we'd like to rewrite (F Int) to Int, getting
1993
1994 case e `cast` co of x'::Int
1995 I# x# -> let x = x' `cast` sym co
1996 in rhs
1997
1998 so that 'rhs' can take advantage of the form of x'.
1999
2000 Notice that Note [Case of cast] (in OccurAnal) may then apply to the result.
2001
2002 Nota Bene: We only do the [Improving seq] transformation if the
2003 case binder 'x' is actually used in the rhs; that is, if the case
2004 is *not* a *pure* seq.
2005 a) There is no point in adding the cast to a pure seq.
2006 b) There is a good reason not to: doing so would interfere
2007 with seq rules (Note [Built-in RULES for seq] in MkId).
2008 In particular, this [Improving seq] thing *adds* a cast
2009 while [Built-in RULES for seq] *removes* one, so they
2010 just flip-flop.
2011
2012 You might worry about
2013 case v of x { __DEFAULT ->
2014 ... case (v `cast` co) of y { I# -> ... }}
2015 This is a pure seq (since x is unused), so [Improving seq] won't happen.
2016 But it's ok: the simplifier will replace 'v' by 'x' in the rhs to get
2017 case v of x { __DEFAULT ->
2018 ... case (x `cast` co) of y { I# -> ... }}
2019 Now the outer case is not a pure seq, so [Improving seq] will happen,
2020 and then the inner case will disappear.
2021
2022 The need for [Improving seq] showed up in Roman's experiments. Example:
2023 foo :: F Int -> Int -> Int
2024 foo t n = t `seq` bar n
2025 where
2026 bar 0 = 0
2027 bar n = bar (n - case t of TI i -> i)
2028 Here we'd like to avoid repeated evaluating t inside the loop, by
2029 taking advantage of the `seq`.
2030
2031 At one point I did transformation in LiberateCase, but it's more
2032 robust here. (Otherwise, there's a danger that we'll simply drop the
2033 'seq' altogether, before LiberateCase gets to see it.)
2034 -}
2035
2036 simplAlts :: SimplEnv
2037 -> OutExpr
2038 -> InId -- Case binder
2039 -> [InAlt] -- Non-empty
2040 -> SimplCont
2041 -> SimplM (OutExpr, OutId, [OutAlt]) -- Includes the continuation
2042 -- Like simplExpr, this just returns the simplified alternatives;
2043 -- it does not return an environment
2044 -- The returned alternatives can be empty, none are possible
2045
2046 simplAlts env scrut case_bndr alts cont'
2047 = do { let env0 = zapFloats env
2048
2049 ; (env1, case_bndr1) <- simplBinder env0 case_bndr
2050 ; let case_bndr2 = case_bndr1 `setIdUnfolding` evaldUnfolding
2051 env2 = modifyInScope env1 case_bndr2
2052 -- See Note [Case-binder evaluated-ness]
2053
2054 ; fam_envs <- getFamEnvs
2055 ; (alt_env', scrut', case_bndr') <- improveSeq fam_envs env2 scrut
2056 case_bndr case_bndr2 alts
2057
2058 ; (imposs_deflt_cons, in_alts) <- prepareAlts scrut' case_bndr' alts
2059 -- NB: it's possible that the returned in_alts is empty: this is handled
2060 -- by the caller (rebuildCase) in the missingAlt function
2061
2062 ; alts' <- mapM (simplAlt alt_env' (Just scrut') imposs_deflt_cons case_bndr' cont') in_alts
2063 ; -- pprTrace "simplAlts" (ppr case_bndr $$ ppr alts_ty $$ ppr alts_ty' $$ ppr alts $$ ppr cont') $
2064 return (scrut', case_bndr', alts') }
2065
2066
2067 ------------------------------------
2068 improveSeq :: (FamInstEnv, FamInstEnv) -> SimplEnv
2069 -> OutExpr -> InId -> OutId -> [InAlt]
2070 -> SimplM (SimplEnv, OutExpr, OutId)
2071 -- Note [Improving seq]
2072 improveSeq fam_envs env scrut case_bndr case_bndr1 [(DEFAULT,_,_)]
2073 | not (isDeadBinder case_bndr) -- Not a pure seq! See Note [Improving seq]
2074 , Just (co, ty2) <- topNormaliseType_maybe fam_envs (idType case_bndr1)
2075 = do { case_bndr2 <- newId (fsLit "nt") ty2
2076 ; let rhs = DoneEx (Var case_bndr2 `Cast` mkSymCo co)
2077 env2 = extendIdSubst env case_bndr rhs
2078 ; return (env2, scrut `Cast` co, case_bndr2) }
2079
2080 improveSeq _ env scrut _ case_bndr1 _
2081 = return (env, scrut, case_bndr1)
2082
2083
2084 ------------------------------------
2085 simplAlt :: SimplEnv
2086 -> Maybe OutExpr -- The scrutinee
2087 -> [AltCon] -- These constructors can't be present when
2088 -- matching the DEFAULT alternative
2089 -> OutId -- The case binder
2090 -> SimplCont
2091 -> InAlt
2092 -> SimplM OutAlt
2093
2094 simplAlt env _ imposs_deflt_cons case_bndr' cont' (DEFAULT, bndrs, rhs)
2095 = ASSERT( null bndrs )
2096 do { let env' = addBinderUnfolding env case_bndr'
2097 (mkOtherCon imposs_deflt_cons)
2098 -- Record the constructors that the case-binder *can't* be.
2099 ; rhs' <- simplExprC env' rhs cont'
2100 ; return (DEFAULT, [], rhs') }
2101
2102 simplAlt env scrut' _ case_bndr' cont' (LitAlt lit, bndrs, rhs)
2103 = ASSERT( null bndrs )
2104 do { env' <- addAltUnfoldings env scrut' case_bndr' (Lit lit)
2105 ; rhs' <- simplExprC env' rhs cont'
2106 ; return (LitAlt lit, [], rhs') }
2107
2108 simplAlt env scrut' _ case_bndr' cont' (DataAlt con, vs, rhs)
2109 = do { -- Deal with the pattern-bound variables
2110 -- Mark the ones that are in ! positions in the
2111 -- data constructor as certainly-evaluated.
2112 -- NB: simplLamBinders preserves this eval info
2113 ; let vs_with_evals = add_evals (dataConRepStrictness con)
2114 ; (env', vs') <- simplLamBndrs env vs_with_evals
2115
2116 -- Bind the case-binder to (con args)
2117 ; let inst_tys' = tyConAppArgs (idType case_bndr')
2118 con_app :: OutExpr
2119 con_app = mkConApp2 con inst_tys' vs'
2120
2121 ; env'' <- addAltUnfoldings env' scrut' case_bndr' con_app
2122 ; rhs' <- simplExprC env'' rhs cont'
2123 ; return (DataAlt con, vs', rhs') }
2124 where
2125 -- add_evals records the evaluated-ness of the bound variables of
2126 -- a case pattern. This is *important*. Consider
2127 -- data T = T !Int !Int
2128 --
2129 -- case x of { T a b -> T (a+1) b }
2130 --
2131 -- We really must record that b is already evaluated so that we don't
2132 -- go and re-evaluate it when constructing the result.
2133 -- See Note [Data-con worker strictness] in MkId.hs
2134 add_evals the_strs
2135 = go vs the_strs
2136 where
2137 go [] [] = []
2138 go (v:vs') strs | isTyVar v = v : go vs' strs
2139 go (v:vs') (str:strs)
2140 | isMarkedStrict str = eval v : go vs' strs
2141 | otherwise = zap v : go vs' strs
2142 go _ _ = pprPanic "cat_evals"
2143 (ppr con $$
2144 ppr vs $$
2145 ppr_with_length the_strs $$
2146 ppr_with_length (dataConRepArgTys con) $$
2147 ppr_with_length (dataConRepStrictness con))
2148 where
2149 ppr_with_length list
2150 = ppr list <+> parens (text "length =" <+> ppr (length list))
2151 -- NB: If this panic triggers, note that
2152 -- NoStrictnessMark doesn't print!
2153
2154 zap v = zapIdOccInfo v -- See Note [Case alternative occ info]
2155 eval v = zap v `setIdUnfolding` evaldUnfolding
2156
2157 addAltUnfoldings :: SimplEnv -> Maybe OutExpr -> OutId -> OutExpr -> SimplM SimplEnv
2158 addAltUnfoldings env scrut case_bndr con_app
2159 = do { dflags <- getDynFlags
2160 ; let con_app_unf = mkSimpleUnfolding dflags con_app
2161 env1 = addBinderUnfolding env case_bndr con_app_unf
2162
2163 -- See Note [Add unfolding for scrutinee]
2164 env2 = case scrut of
2165 Just (Var v) -> addBinderUnfolding env1 v con_app_unf
2166 Just (Cast (Var v) co) -> addBinderUnfolding env1 v $
2167 mkSimpleUnfolding dflags (Cast con_app (mkSymCo co))
2168 _ -> env1
2169
2170 ; traceSmpl "addAltUnf" (vcat [ppr case_bndr <+> ppr scrut, ppr con_app])
2171 ; return env2 }
2172
2173 addBinderUnfolding :: SimplEnv -> Id -> Unfolding -> SimplEnv
2174 addBinderUnfolding env bndr unf
2175 | debugIsOn, Just tmpl <- maybeUnfoldingTemplate unf
2176 = WARN( not (eqType (idType bndr) (exprType tmpl)),
2177 ppr bndr $$ ppr (idType bndr) $$ ppr tmpl $$ ppr (exprType tmpl) )
2178 modifyInScope env (bndr `setIdUnfolding` unf)
2179
2180 | otherwise
2181 = modifyInScope env (bndr `setIdUnfolding` unf)
2182
2183 zapBndrOccInfo :: Bool -> Id -> Id
2184 -- Consider case e of b { (a,b) -> ... }
2185 -- Then if we bind b to (a,b) in "...", and b is not dead,
2186 -- then we must zap the deadness info on a,b
2187 zapBndrOccInfo keep_occ_info pat_id
2188 | keep_occ_info = pat_id
2189 | otherwise = zapIdOccInfo pat_id
2190
2191 {- Note [Case binder evaluated-ness]
2192 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2193 We pin on a (OtherCon []) unfolding to the case-binder of a Case,
2194 even though it'll be over-ridden in every case alternative with a more
2195 informative unfolding. Why? Because suppose a later, less clever, pass
2196 simply replaces all occurrences of the case binder with the binder itself;
2197 then Lint may complain about the let/app invariant. Example
2198 case e of b { DEFAULT -> let v = reallyUnsafePtrEq# b y in ....
2199 ; K -> blah }
2200
2201 The let/app invariant requires that y is evaluated in the call to
2202 reallyUnsafePtrEq#, which it is. But we still want that to be true if we
2203 propagate binders to occurrences.
2204
2205 This showed up in Trac #13027.
2206
2207 Note [Add unfolding for scrutinee]
2208 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2209 In general it's unlikely that a variable scrutinee will appear
2210 in the case alternatives case x of { ...x unlikely to appear... }
2211 because the binder-swap in OccAnal has got rid of all such occcurrences
2212 See Note [Binder swap] in OccAnal.
2213
2214 BUT it is still VERY IMPORTANT to add a suitable unfolding for a
2215 variable scrutinee, in simplAlt. Here's why
2216 case x of y
2217 (a,b) -> case b of c
2218 I# v -> ...(f y)...
2219 There is no occurrence of 'b' in the (...(f y)...). But y gets
2220 the unfolding (a,b), and *that* mentions b. If f has a RULE
2221 RULE f (p, I# q) = ...
2222 we want that rule to match, so we must extend the in-scope env with a
2223 suitable unfolding for 'y'. It's *essential* for rule matching; but
2224 it's also good for case-elimintation -- suppose that 'f' was inlined
2225 and did multi-level case analysis, then we'd solve it in one
2226 simplifier sweep instead of two.
2227
2228 Exactly the same issue arises in SpecConstr;
2229 see Note [Add scrutinee to ValueEnv too] in SpecConstr
2230
2231 HOWEVER, given
2232 case x of y { Just a -> r1; Nothing -> r2 }
2233 we do not want to add the unfolding x -> y to 'x', which might seem cool,
2234 since 'y' itself has different unfoldings in r1 and r2. Reason: if we
2235 did that, we'd have to zap y's deadness info and that is a very useful
2236 piece of information.
2237
2238 So instead we add the unfolding x -> Just a, and x -> Nothing in the
2239 respective RHSs.
2240
2241
2242 ************************************************************************
2243 * *
2244 \subsection{Known constructor}
2245 * *
2246 ************************************************************************
2247
2248 We are a bit careful with occurrence info. Here's an example
2249
2250 (\x* -> case x of (a*, b) -> f a) (h v, e)
2251
2252 where the * means "occurs once". This effectively becomes
2253 case (h v, e) of (a*, b) -> f a)
2254 and then
2255 let a* = h v; b = e in f a
2256 and then
2257 f (h v)
2258
2259 All this should happen in one sweep.
2260 -}
2261
2262 knownCon :: SimplEnv
2263 -> OutExpr -- The scrutinee
2264 -> DataCon -> [OutType] -> [OutExpr] -- The scrutinee (in pieces)
2265 -> InId -> [InBndr] -> InExpr -- The alternative
2266 -> SimplCont
2267 -> SimplM (SimplEnv, OutExpr)
2268
2269 knownCon env scrut dc dc_ty_args dc_args bndr bs rhs cont
2270 = do { env' <- bind_args env bs dc_args
2271 ; env'' <- bind_case_bndr env'
2272 ; simplExprF env'' rhs cont }
2273 where
2274 zap_occ = zapBndrOccInfo (isDeadBinder bndr) -- bndr is an InId
2275
2276 -- Ugh!
2277 bind_args env' [] _ = return env'
2278
2279 bind_args env' (b:bs') (Type ty : args)
2280 = ASSERT( isTyVar b )
2281 bind_args (extendTvSubst env' b ty) bs' args
2282
2283 bind_args env' (b:bs') (Coercion co : args)
2284 = ASSERT( isCoVar b )
2285 bind_args (extendCvSubst env' b co) bs' args
2286
2287 bind_args env' (b:bs') (arg : args)
2288 = ASSERT( isId b )
2289 do { let b' = zap_occ b
2290 -- Note that the binder might be "dead", because it doesn't
2291 -- occur in the RHS; and simplNonRecX may therefore discard
2292 -- it via postInlineUnconditionally.
2293 -- Nevertheless we must keep it if the case-binder is alive,
2294 -- because it may be used in the con_app. See Note [knownCon occ info]
2295 ; env'' <- simplNonRecX env' b' arg -- arg satisfies let/app invariant
2296 ; bind_args env'' bs' args }
2297
2298 bind_args _ _ _ =
2299 pprPanic "bind_args" $ ppr dc $$ ppr bs $$ ppr dc_args $$
2300 text "scrut:" <+> ppr scrut
2301
2302 -- It's useful to bind bndr to scrut, rather than to a fresh
2303 -- binding x = Con arg1 .. argn
2304 -- because very often the scrut is a variable, so we avoid
2305 -- creating, and then subsequently eliminating, a let-binding
2306 -- BUT, if scrut is a not a variable, we must be careful
2307 -- about duplicating the arg redexes; in that case, make
2308 -- a new con-app from the args
2309 bind_case_bndr env
2310 | isDeadBinder bndr = return env
2311 | exprIsTrivial scrut = return (extendIdSubst env bndr (DoneEx scrut))
2312 | otherwise = do { dc_args <- mapM (simplVar env) bs
2313 -- dc_ty_args are aready OutTypes,
2314 -- but bs are InBndrs
2315 ; let con_app = Var (dataConWorkId dc)
2316 `mkTyApps` dc_ty_args
2317 `mkApps` dc_args
2318 ; simplNonRecX env bndr con_app }
2319
2320 -------------------
2321 missingAlt :: SimplEnv -> Id -> [InAlt] -> SimplCont -> SimplM (SimplEnv, OutExpr)
2322 -- This isn't strictly an error, although it is unusual.
2323 -- It's possible that the simplifier might "see" that
2324 -- an inner case has no accessible alternatives before
2325 -- it "sees" that the entire branch of an outer case is
2326 -- inaccessible. So we simply put an error case here instead.
2327 missingAlt env case_bndr _ cont
2328 = WARN( True, text "missingAlt" <+> ppr case_bndr )
2329 return (env, mkImpossibleExpr (contResultType cont))
2330
2331 {-
2332 ************************************************************************
2333 * *
2334 \subsection{Duplicating continuations}
2335 * *
2336 ************************************************************************
2337 -}
2338
2339 prepareCaseCont :: SimplEnv
2340 -> [InAlt] -> SimplCont
2341 -> SimplM (SimplEnv,
2342 SimplCont, -- Dupable part
2343 SimplCont) -- Non-dupable part
2344 -- We are considering
2345 -- K[case _ of { p1 -> r1; ...; pn -> rn }]
2346 -- where K is some enclosing continuation for the case
2347 -- Goal: split K into two pieces Kdup,Knodup so that
2348 -- a) Kdup can be duplicated
2349 -- b) Knodup[Kdup[e]] = K[e]
2350 -- The idea is that we'll transform thus:
2351 -- Knodup[ (case _ of { p1 -> Kdup[r1]; ...; pn -> Kdup[rn] }
2352 --
2353 -- We may also return some extra bindings in SimplEnv (that scope over
2354 -- the entire continuation)
2355 --
2356 -- When case-of-case is off, just make the entire continuation non-dupable
2357
2358 prepareCaseCont env alts cont
2359 | not (sm_case_case (getMode env)) = return (env, mkBoringStop (contHoleType cont), cont)
2360 | not (many_alts alts) = return (env, cont, mkBoringStop (contResultType cont))
2361 | otherwise = mkDupableCont env cont
2362 where
2363 many_alts :: [InAlt] -> Bool -- True iff strictly > 1 non-bottom alternative
2364 many_alts [] = False -- See Note [Bottom alternatives]
2365 many_alts [_] = False
2366 many_alts (alt:alts)
2367 | is_bot_alt alt = many_alts alts
2368 | otherwise = not (all is_bot_alt alts)
2369
2370 is_bot_alt (_,_,rhs) = exprIsBottom rhs
2371
2372 {-
2373 Note [Bottom alternatives]
2374 ~~~~~~~~~~~~~~~~~~~~~~~~~~
2375 When we have
2376 case (case x of { A -> error .. ; B -> e; C -> error ..)
2377 of alts
2378 then we can just duplicate those alts because the A and C cases
2379 will disappear immediately. This is more direct than creating
2380 join points and inlining them away; and in some cases we would
2381 not even create the join points (see Note [Single-alternative case])
2382 and we would keep the case-of-case which is silly. See Trac #4930.
2383 -}
2384
2385 mkDupableCont :: SimplEnv -> SimplCont
2386 -> SimplM (SimplEnv, SimplCont, SimplCont)
2387
2388 mkDupableCont env cont
2389 | contIsDupable cont
2390 = return (env, cont, mkBoringStop (contResultType cont))
2391
2392 mkDupableCont _ (Stop {}) = panic "mkDupableCont" -- Handled by previous eqn
2393
2394 mkDupableCont env (CastIt ty cont)
2395 = do { (env', dup, nodup) <- mkDupableCont env cont
2396 ; return (env', CastIt ty dup, nodup) }
2397
2398 -- Duplicating ticks for now, not sure if this is good or not
2399 mkDupableCont env cont@(TickIt{})
2400 = return (env, mkBoringStop (contHoleType cont), cont)
2401
2402 mkDupableCont env cont@(StrictBind {})
2403 = return (env, mkBoringStop (contHoleType cont), cont)
2404 -- See Note [Duplicating StrictBind]
2405
2406 mkDupableCont env (StrictArg info cci cont)
2407 -- See Note [Duplicating StrictArg]
2408 = do { (env', dup, nodup) <- mkDupableCont env cont
2409 ; (env'', args') <- mapAccumLM makeTrivialArg env' (ai_args info)
2410 ; return (env'', StrictArg (info { ai_args = args' }) cci dup, nodup) }
2411
2412 mkDupableCont env cont@(ApplyToTy { sc_cont = tail })
2413 = do { (env', dup_cont, nodup_cont) <- mkDupableCont env tail
2414 ; return (env', cont { sc_cont = dup_cont }, nodup_cont ) }
2415
2416 mkDupableCont env (ApplyToVal { sc_arg = arg, sc_dup = dup, sc_env = se, sc_cont = cont })
2417 = -- e.g. [...hole...] (...arg...)
2418 -- ==>
2419 -- let a = ...arg...
2420 -- in [...hole...] a
2421 do { (env', dup_cont, nodup_cont) <- mkDupableCont env cont
2422 ; (_, se', arg') <- simplArg env' dup se arg
2423 ; (env'', arg'') <- makeTrivial NotTopLevel env' (fsLit "karg") arg'
2424 ; let app_cont = ApplyToVal { sc_arg = arg'', sc_env = se'
2425 , sc_dup = OkToDup, sc_cont = dup_cont }
2426 ; return (env'', app_cont, nodup_cont) }
2427
2428 mkDupableCont env cont@(Select { sc_bndr = case_bndr, sc_alts = [(_, bs, _rhs)] })
2429 -- See Note [Single-alternative case]
2430 -- | not (exprIsDupable rhs && contIsDupable case_cont)
2431 -- | not (isDeadBinder case_bndr)
2432 | all isDeadBinder bs -- InIds
2433 && not (isUnliftedType (idType case_bndr))
2434 -- Note [Single-alternative-unlifted]
2435 = return (env, mkBoringStop (contHoleType cont), cont)
2436
2437 mkDupableCont env (Select { sc_bndr = case_bndr, sc_alts = alts
2438 , sc_env = se, sc_cont = cont })
2439 = -- e.g. (case [...hole...] of { pi -> ei })
2440 -- ===>
2441 -- let ji = \xij -> ei
2442 -- in case [...hole...] of { pi -> ji xij }
2443 do { tick (CaseOfCase case_bndr)
2444 ; (env', dup_cont, nodup_cont) <- prepareCaseCont env alts cont
2445 -- NB: We call prepareCaseCont here. If there is only one
2446 -- alternative, then dup_cont may be big, but that's ok
2447 -- because we push it into the single alternative, and then
2448 -- use mkDupableAlt to turn that simplified alternative into
2449 -- a join point if it's too big to duplicate.
2450 -- And this is important: see Note [Fusing case continuations]
2451
2452 ; let alt_env = se `setInScope` env'
2453
2454 ; (alt_env', case_bndr') <- simplBinder alt_env case_bndr
2455 ; alts' <- mapM (simplAlt alt_env' Nothing [] case_bndr' dup_cont) alts
2456 -- Safe to say that there are no handled-cons for the DEFAULT case
2457 -- NB: simplBinder does not zap deadness occ-info, so
2458 -- a dead case_bndr' will still advertise its deadness
2459 -- This is really important because in
2460 -- case e of b { (# p,q #) -> ... }
2461 -- b is always dead, and indeed we are not allowed to bind b to (# p,q #),
2462 -- which might happen if e was an explicit unboxed pair and b wasn't marked dead.
2463 -- In the new alts we build, we have the new case binder, so it must retain
2464 -- its deadness.
2465 -- NB: we don't use alt_env further; it has the substEnv for
2466 -- the alternatives, and we don't want that
2467
2468 ; (env'', alts'') <- mkDupableAlts env' case_bndr' alts'
2469 ; return (env'', -- Note [Duplicated env]
2470 Select { sc_dup = OkToDup
2471 , sc_bndr = case_bndr', sc_alts = alts''
2472 , sc_env = zapSubstEnv env''
2473 , sc_cont = mkBoringStop (contHoleType nodup_cont) },
2474 nodup_cont) }
2475
2476
2477 mkDupableAlts :: SimplEnv -> OutId -> [InAlt]
2478 -> SimplM (SimplEnv, [InAlt])
2479 -- Absorbs the continuation into the new alternatives
2480
2481 mkDupableAlts env case_bndr' the_alts
2482 = go env the_alts
2483 where
2484 go env0 [] = return (env0, [])
2485 go env0 (alt:alts)
2486 = do { (env1, alt') <- mkDupableAlt env0 case_bndr' alt
2487 ; (env2, alts') <- go env1 alts
2488 ; return (env2, alt' : alts' ) }
2489
2490 mkDupableAlt :: SimplEnv -> OutId -> (AltCon, [CoreBndr], CoreExpr)
2491 -> SimplM (SimplEnv, (AltCon, [CoreBndr], CoreExpr))
2492 mkDupableAlt env case_bndr (con, bndrs', rhs') = do
2493 dflags <- getDynFlags
2494 if exprIsDupable dflags rhs' -- Note [Small alternative rhs]
2495 then return (env, (con, bndrs', rhs'))
2496 else
2497 do { let rhs_ty' = exprType rhs'
2498 scrut_ty = idType case_bndr
2499 case_bndr_w_unf
2500 = case con of
2501 DEFAULT -> case_bndr
2502 DataAlt dc -> setIdUnfolding case_bndr unf
2503 where
2504 -- See Note [Case binders and join points]
2505 unf = mkInlineUnfolding rhs
2506 rhs = mkConApp2 dc (tyConAppArgs scrut_ty) bndrs'
2507
2508 LitAlt {} -> WARN( True, text "mkDupableAlt"
2509 <+> ppr case_bndr <+> ppr con )
2510 case_bndr
2511 -- The case binder is alive but trivial, so why has
2512 -- it not been substituted away?
2513
2514 used_bndrs' | isDeadBinder case_bndr = filter abstract_over bndrs'
2515 | otherwise = bndrs' ++ [case_bndr_w_unf]
2516
2517 abstract_over bndr
2518 | isTyVar bndr = True -- Abstract over all type variables just in case
2519 | otherwise = not (isDeadBinder bndr)
2520 -- The deadness info on the new Ids is preserved by simplBinders
2521
2522 ; (final_bndrs', final_args) -- Note [Join point abstraction]
2523 <- if (any isId used_bndrs')
2524 then return (used_bndrs', varsToCoreExprs used_bndrs')
2525 else do { rw_id <- newId (fsLit "w") voidPrimTy
2526 ; return ([setOneShotLambda rw_id], [Var voidPrimId]) }
2527
2528 ; join_bndr <- newId (fsLit "$j") (mkLamTypes final_bndrs' rhs_ty')
2529 -- Note [Funky mkLamTypes]
2530
2531 ; let -- We make the lambdas into one-shot-lambdas. The
2532 -- join point is sure to be applied at most once, and doing so
2533 -- prevents the body of the join point being floated out by
2534 -- the full laziness pass
2535 really_final_bndrs = map one_shot final_bndrs'
2536 one_shot v | isId v = setOneShotLambda v
2537 | otherwise = v
2538 join_rhs = mkLams really_final_bndrs rhs'
2539 join_arity = exprArity join_rhs
2540 join_call = mkApps (Var join_bndr) final_args
2541
2542 ; env' <- addPolyBind NotTopLevel env (NonRec (join_bndr `setIdArity` join_arity) join_rhs)
2543 ; return (env', (con, bndrs', join_call)) }
2544 -- See Note [Duplicated env]
2545
2546 {-
2547 Note [Fusing case continuations]
2548 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2549 It's important to fuse two successive case continuations when the
2550 first has one alternative. That's why we call prepareCaseCont here.
2551 Consider this, which arises from thunk splitting (see Note [Thunk
2552 splitting] in WorkWrap):
2553
2554 let
2555 x* = case (case v of {pn -> rn}) of
2556 I# a -> I# a
2557 in body
2558
2559 The simplifier will find
2560 (Var v) with continuation
2561 Select (pn -> rn) (
2562 Select [I# a -> I# a] (
2563 StrictBind body Stop
2564
2565 So we'll call mkDupableCont on
2566 Select [I# a -> I# a] (StrictBind body Stop)
2567 There is just one alternative in the first Select, so we want to
2568 simplify the rhs (I# a) with continuation (StricgtBind body Stop)
2569 Supposing that body is big, we end up with
2570 let $j a = <let x = I# a in body>
2571 in case v of { pn -> case rn of
2572 I# a -> $j a }
2573 This is just what we want because the rn produces a box that
2574 the case rn cancels with.
2575
2576 See Trac #4957 a fuller example.
2577
2578 Note [Case binders and join points]
2579 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2580 Consider this
2581 case (case .. ) of c {
2582 I# c# -> ....c....
2583
2584 If we make a join point with c but not c# we get
2585 $j = \c -> ....c....
2586
2587 But if later inlining scrutinises the c, thus
2588
2589 $j = \c -> ... case c of { I# y -> ... } ...
2590
2591 we won't see that 'c' has already been scrutinised. This actually
2592 happens in the 'tabulate' function in wave4main, and makes a significant
2593 difference to allocation.
2594
2595 An alternative plan is this:
2596
2597 $j = \c# -> let c = I# c# in ...c....
2598
2599 but that is bad if 'c' is *not* later scrutinised.
2600
2601 So instead we do both: we pass 'c' and 'c#' , and record in c's inlining
2602 (a stable unfolding) that it's really I# c#, thus
2603
2604 $j = \c# -> \c[=I# c#] -> ...c....
2605
2606 Absence analysis may later discard 'c'.
2607
2608 NB: take great care when doing strictness analysis;
2609 see Note [Lamba-bound unfoldings] in DmdAnal.
2610
2611 Also note that we can still end up passing stuff that isn't used. Before
2612 strictness analysis we have
2613 let $j x y c{=(x,y)} = (h c, ...)
2614 in ...
2615 After strictness analysis we see that h is strict, we end up with
2616 let $j x y c{=(x,y)} = ($wh x y, ...)
2617 and c is unused.
2618
2619 Note [Duplicated env]
2620 ~~~~~~~~~~~~~~~~~~~~~
2621 Some of the alternatives are simplified, but have not been turned into a join point
2622 So they *must* have an zapped subst-env. So we can't use completeNonRecX to
2623 bind the join point, because it might to do PostInlineUnconditionally, and
2624 we'd lose that when zapping the subst-env. We could have a per-alt subst-env,
2625 but zapping it (as we do in mkDupableCont, the Select case) is safe, and
2626 at worst delays the join-point inlining.
2627
2628 Note [Small alternative rhs]
2629 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2630 It is worth checking for a small RHS because otherwise we
2631 get extra let bindings that may cause an extra iteration of the simplifier to
2632 inline back in place. Quite often the rhs is just a variable or constructor.
2633 The Ord instance of Maybe in PrelMaybe.hs, for example, took several extra
2634 iterations because the version with the let bindings looked big, and so wasn't
2635 inlined, but after the join points had been inlined it looked smaller, and so
2636 was inlined.
2637
2638 NB: we have to check the size of rhs', not rhs.
2639 Duplicating a small InAlt might invalidate occurrence information
2640 However, if it *is* dupable, we return the *un* simplified alternative,
2641 because otherwise we'd need to pair it up with an empty subst-env....
2642 but we only have one env shared between all the alts.
2643 (Remember we must zap the subst-env before re-simplifying something).
2644 Rather than do this we simply agree to re-simplify the original (small) thing later.
2645
2646 Note [Funky mkLamTypes]
2647 ~~~~~~~~~~~~~~~~~~~~~~
2648 Notice the funky mkLamTypes. If the constructor has existentials
2649 it's possible that the join point will be abstracted over
2650 type variables as well as term variables.
2651 Example: Suppose we have
2652 data T = forall t. C [t]
2653 Then faced with
2654 case (case e of ...) of
2655 C t xs::[t] -> rhs
2656 We get the join point
2657 let j :: forall t. [t] -> ...
2658 j = /\t \xs::[t] -> rhs
2659 in
2660 case (case e of ...) of
2661 C t xs::[t] -> j t xs
2662
2663 Note [Join point abstraction]
2664 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2665 Join points always have at least one value argument,
2666 for several reasons
2667
2668 * If we try to lift a primitive-typed something out
2669 for let-binding-purposes, we will *caseify* it (!),
2670 with potentially-disastrous strictness results. So
2671 instead we turn it into a function: \v -> e
2672 where v::Void#. The value passed to this function is void,
2673 which generates (almost) no code.
2674
2675 * CPR. We used to say "&& isUnliftedType rhs_ty'" here, but now
2676 we make the join point into a function whenever used_bndrs'
2677 is empty. This makes the join-point more CPR friendly.
2678 Consider: let j = if .. then I# 3 else I# 4
2679 in case .. of { A -> j; B -> j; C -> ... }
2680
2681 Now CPR doesn't w/w j because it's a thunk, so
2682 that means that the enclosing function can't w/w either,
2683 which is a lose. Here's the example that happened in practice:
2684 kgmod :: Int -> Int -> Int
2685 kgmod x y = if x > 0 && y < 0 || x < 0 && y > 0
2686 then 78
2687 else 5
2688
2689 * Let-no-escape. We want a join point to turn into a let-no-escape
2690 so that it is implemented as a jump, and one of the conditions
2691 for LNE is that it's not updatable. In CoreToStg, see
2692 Note [What is a non-escaping let]
2693
2694 * Floating. Since a join point will be entered once, no sharing is
2695 gained by floating out, but something might be lost by doing
2696 so because it might be allocated.
2697
2698 I have seen a case alternative like this:
2699 True -> \v -> ...
2700 It's a bit silly to add the realWorld dummy arg in this case, making
2701 $j = \s v -> ...
2702 True -> $j s
2703 (the \v alone is enough to make CPR happy) but I think it's rare
2704
2705 There's a slight infelicity here: we pass the overall
2706 case_bndr to all the join points if it's used in *any* RHS,
2707 because we don't know its usage in each RHS separately
2708
2709
2710 Note [Duplicating StrictArg]
2711 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2712 The original plan had (where E is a big argument)
2713 e.g. f E [..hole..]
2714 ==> let $j = \a -> f E a
2715 in $j [..hole..]
2716
2717 But this is terrible! Here's an example:
2718 && E (case x of { T -> F; F -> T })
2719 Now, && is strict so we end up simplifying the case with
2720
2721 an ArgOf continuation. If we let-bind it, we get
2722 let $j = \v -> && E v
2723 in simplExpr (case x of { T -> F; F -> T })
2724 (ArgOf (\r -> $j r)
2725 And after simplifying more we get
2726 let $j = \v -> && E v
2727 in case x of { T -> $j F; F -> $j T }
2728 Which is a Very Bad Thing
2729
2730 What we do now is this
2731 f E [..hole..]
2732 ==> let a = E
2733 in f a [..hole..]
2734 Now if the thing in the hole is a case expression (which is when
2735 we'll call mkDupableCont), we'll push the function call into the
2736 branches, which is what we want. Now RULES for f may fire, and
2737 call-pattern specialisation. Here's an example from Trac #3116
2738 go (n+1) (case l of
2739 1 -> bs'
2740 _ -> Chunk p fpc (o+1) (l-1) bs')
2741 If we can push the call for 'go' inside the case, we get
2742 call-pattern specialisation for 'go', which is *crucial* for
2743 this program.
2744
2745 Here is the (&&) example:
2746 && E (case x of { T -> F; F -> T })
2747 ==> let a = E in
2748 case x of { T -> && a F; F -> && a T }
2749 Much better!
2750
2751 Notice that
2752 * Arguments to f *after* the strict one are handled by
2753 the ApplyToVal case of mkDupableCont. Eg
2754 f [..hole..] E
2755
2756 * We can only do the let-binding of E because the function
2757 part of a StrictArg continuation is an explicit syntax
2758 tree. In earlier versions we represented it as a function
2759 (CoreExpr -> CoreEpxr) which we couldn't take apart.
2760
2761 Do *not* duplicate StrictBind and StritArg continuations. We gain
2762 nothing by propagating them into the expressions, and we do lose a
2763 lot.
2764
2765 The desire not to duplicate is the entire reason that
2766 mkDupableCont returns a pair of continuations.
2767
2768 Note [Duplicating StrictBind]
2769 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2770 Unlike StrictArg, there doesn't seem anything to gain from
2771 duplicating a StrictBind continuation, so we don't.
2772
2773
2774 Note [Single-alternative cases]
2775 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2776 This case is just like the ArgOf case. Here's an example:
2777 data T a = MkT !a
2778 ...(MkT (abs x))...
2779 Then we get
2780 case (case x of I# x' ->
2781 case x' <# 0# of
2782 True -> I# (negate# x')
2783 False -> I# x') of y {
2784 DEFAULT -> MkT y
2785 Because the (case x) has only one alternative, we'll transform to
2786 case x of I# x' ->
2787 case (case x' <# 0# of
2788 True -> I# (negate# x')
2789 False -> I# x') of y {
2790 DEFAULT -> MkT y
2791 But now we do *NOT* want to make a join point etc, giving
2792 case x of I# x' ->
2793 let $j = \y -> MkT y
2794 in case x' <# 0# of
2795 True -> $j (I# (negate# x'))
2796 False -> $j (I# x')
2797 In this case the $j will inline again, but suppose there was a big
2798 strict computation enclosing the orginal call to MkT. Then, it won't
2799 "see" the MkT any more, because it's big and won't get duplicated.
2800 And, what is worse, nothing was gained by the case-of-case transform.
2801
2802 So, in circumstances like these, we don't want to build join points
2803 and push the outer case into the branches of the inner one. Instead,
2804 don't duplicate the continuation.
2805
2806 When should we use this strategy? We should not use it on *every*
2807 single-alternative case:
2808 e.g. case (case ....) of (a,b) -> (# a,b #)
2809 Here we must push the outer case into the inner one!
2810 Other choices:
2811
2812 * Match [(DEFAULT,_,_)], but in the common case of Int,
2813 the alternative-filling-in code turned the outer case into
2814 case (...) of y { I# _ -> MkT y }
2815
2816 * Match on single alternative plus (not (isDeadBinder case_bndr))
2817 Rationale: pushing the case inwards won't eliminate the construction.
2818 But there's a risk of
2819 case (...) of y { (a,b) -> let z=(a,b) in ... }
2820 Now y looks dead, but it'll come alive again. Still, this
2821 seems like the best option at the moment.
2822
2823 * Match on single alternative plus (all (isDeadBinder bndrs))
2824 Rationale: this is essentially seq.
2825
2826 * Match when the rhs is *not* duplicable, and hence would lead to a
2827 join point. This catches the disaster-case above. We can test
2828 the *un-simplified* rhs, which is fine. It might get bigger or
2829 smaller after simplification; if it gets smaller, this case might
2830 fire next time round. NB also that we must test contIsDupable
2831 case_cont *too, because case_cont might be big!
2832
2833 HOWEVER: I found that this version doesn't work well, because
2834 we can get let x = case (...) of { small } in ...case x...
2835 When x is inlined into its full context, we find that it was a bad
2836 idea to have pushed the outer case inside the (...) case.
2837
2838 There is a cost to not doing case-of-case; see Trac #10626.
2839
2840 Note [Single-alternative-unlifted]
2841 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2842 Here's another single-alternative where we really want to do case-of-case:
2843
2844 data Mk1 = Mk1 Int# | Mk2 Int#
2845
2846 M1.f =
2847 \r [x_s74 y_s6X]
2848 case
2849 case y_s6X of tpl_s7m {
2850 M1.Mk1 ipv_s70 -> ipv_s70;
2851 M1.Mk2 ipv_s72 -> ipv_s72;
2852 }
2853 of
2854 wild_s7c
2855 { __DEFAULT ->
2856 case
2857 case x_s74 of tpl_s7n {
2858 M1.Mk1 ipv_s77 -> ipv_s77;
2859 M1.Mk2 ipv_s79 -> ipv_s79;
2860 }
2861 of
2862 wild1_s7b
2863 { __DEFAULT -> ==# [wild1_s7b wild_s7c];
2864 };
2865 };
2866
2867 So the outer case is doing *nothing at all*, other than serving as a
2868 join-point. In this case we really want to do case-of-case and decide
2869 whether to use a real join point or just duplicate the continuation:
2870
2871 let $j s7c = case x of
2872 Mk1 ipv77 -> (==) s7c ipv77
2873 Mk1 ipv79 -> (==) s7c ipv79
2874 in
2875 case y of
2876 Mk1 ipv70 -> $j ipv70
2877 Mk2 ipv72 -> $j ipv72
2878
2879 Hence: check whether the case binder's type is unlifted, because then
2880 the outer case is *not* a seq.
2881
2882 ************************************************************************
2883 * *
2884 Unfoldings
2885 * *
2886 ************************************************************************
2887 -}
2888
2889 simplLetUnfolding :: SimplEnv-> TopLevelFlag
2890 -> InId
2891 -> OutExpr
2892 -> Unfolding -> SimplM Unfolding
2893 simplLetUnfolding env top_lvl id new_rhs unf
2894 | isStableUnfolding unf
2895 = simplUnfolding env top_lvl id unf
2896 | otherwise
2897 = is_bottoming `seq` -- See Note [Force bottoming field]
2898 do { dflags <- getDynFlags
2899 ; return (mkUnfolding dflags InlineRhs is_top_lvl is_bottoming new_rhs) }
2900 -- We make an unfolding *even for loop-breakers*.
2901 -- Reason: (a) It might be useful to know that they are WHNF
2902 -- (b) In TidyPgm we currently assume that, if we want to
2903 -- expose the unfolding then indeed we *have* an unfolding
2904 -- to expose. (We could instead use the RHS, but currently
2905 -- we don't.) The simple thing is always to have one.
2906 where
2907 is_top_lvl = isTopLevel top_lvl
2908 is_bottoming = isBottomingId id
2909
2910 simplUnfolding :: SimplEnv-> TopLevelFlag -> InId -> Unfolding -> SimplM Unfolding
2911 -- Note [Setting the new unfolding]
2912 simplUnfolding env top_lvl id unf
2913 = case unf of
2914 NoUnfolding -> return unf
2915 BootUnfolding -> return unf
2916 OtherCon {} -> return unf
2917
2918 DFunUnfolding { df_bndrs = bndrs, df_con = con, df_args = args }
2919 -> do { (env', bndrs') <- simplBinders rule_env bndrs
2920 ; args' <- mapM (simplExpr env') args
2921 ; return (mkDFunUnfolding bndrs' con args') }
2922
2923 CoreUnfolding { uf_tmpl = expr, uf_src = src, uf_guidance = guide }
2924 | isStableSource src
2925 -> do { expr' <- simplExpr rule_env expr
2926 ; case guide of
2927 UnfWhen { ug_arity = arity, ug_unsat_ok = sat_ok } -- Happens for INLINE things
2928 -> let guide' = UnfWhen { ug_arity = arity, ug_unsat_ok = sat_ok
2929 , ug_boring_ok = inlineBoringOk expr' }
2930 -- Refresh the boring-ok flag, in case expr'
2931 -- has got small. This happens, notably in the inlinings
2932 -- for dfuns for single-method classes; see
2933 -- Note [Single-method classes] in TcInstDcls.
2934 -- A test case is Trac #4138
2935 in return (mkCoreUnfolding src is_top_lvl expr' guide')
2936 -- See Note [Top-level flag on inline rules] in CoreUnfold
2937
2938 _other -- Happens for INLINABLE things
2939 -> is_bottoming `seq` -- See Note [Force bottoming field]
2940 do { dflags <- getDynFlags
2941 ; return (mkUnfolding dflags src is_top_lvl is_bottoming expr') } }
2942 -- If the guidance is UnfIfGoodArgs, this is an INLINABLE
2943 -- unfolding, and we need to make sure the guidance is kept up
2944 -- to date with respect to any changes in the unfolding.
2945
2946 | otherwise -> return noUnfolding -- Discard unstable unfoldings
2947 where
2948 is_top_lvl = isTopLevel top_lvl
2949 is_bottoming = isBottomingId id
2950 act = idInlineActivation id
2951 rule_env = updMode (updModeForStableUnfoldings act) env
2952 -- See Note [Simplifying inside stable unfoldings] in SimplUtils
2953
2954 {-
2955 Note [Force bottoming field]
2956 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2957 We need to force bottoming, or the new unfolding holds
2958 on to the old unfolding (which is part of the id).
2959
2960 Note [Setting the new unfolding]
2961 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2962 * If there's an INLINE pragma, we simplify the RHS gently. Maybe we
2963 should do nothing at all, but simplifying gently might get rid of
2964 more crap.
2965
2966 * If not, we make an unfolding from the new RHS. But *only* for
2967 non-loop-breakers. Making loop breakers not have an unfolding at all
2968 means that we can avoid tests in exprIsConApp, for example. This is
2969 important: if exprIsConApp says 'yes' for a recursive thing, then we
2970 can get into an infinite loop
2971
2972 If there's an stable unfolding on a loop breaker (which happens for
2973 INLINABLE), we hang on to the inlining. It's pretty dodgy, but the
2974 user did say 'INLINE'. May need to revisit this choice.
2975
2976 ************************************************************************
2977 * *
2978 Rules
2979 * *
2980 ************************************************************************
2981
2982 Note [Rules in a letrec]
2983 ~~~~~~~~~~~~~~~~~~~~~~~~
2984 After creating fresh binders for the binders of a letrec, we
2985 substitute the RULES and add them back onto the binders; this is done
2986 *before* processing any of the RHSs. This is important. Manuel found
2987 cases where he really, really wanted a RULE for a recursive function
2988 to apply in that function's own right-hand side.
2989
2990 See Note [Loop breaking and RULES] in OccAnal.
2991 -}
2992
2993 addBndrRules :: SimplEnv -> InBndr -> OutBndr -> SimplM (SimplEnv, OutBndr)
2994 -- Rules are added back into the bin
2995 addBndrRules env in_id out_id
2996 | null old_rules
2997 = return (env, out_id)
2998 | otherwise
2999 = do { new_rules <- simplRules env (Just (idName out_id)) old_rules
3000 ; let final_id = out_id `setIdSpecialisation` mkRuleInfo new_rules
3001 ; return (modifyInScope env final_id, final_id) }
3002 where
3003 old_rules = ruleInfoRules (idSpecialisation in_id)
3004
3005 simplRules :: SimplEnv -> Maybe Name -> [CoreRule] -> SimplM [CoreRule]
3006 simplRules env mb_new_nm rules
3007 = mapM simpl_rule rules
3008 where
3009 simpl_rule rule@(BuiltinRule {})
3010 = return rule
3011
3012 simpl_rule rule@(Rule { ru_bndrs = bndrs, ru_args = args
3013 , ru_fn = fn_name, ru_rhs = rhs })
3014 = do { (env', bndrs') <- simplBinders env bndrs
3015 ; let rule_env = updMode updModeForRules env'
3016 ; args' <- mapM (simplExpr rule_env) args
3017 ; rhs' <- simplExpr rule_env rhs
3018 ; return (rule { ru_bndrs = bndrs'
3019 , ru_fn = mb_new_nm `orElse` fn_name
3020 , ru_args = args'
3021 , ru_rhs = rhs' }) }